# EAGLE4 Architecture Book (agent view)

Plain-text twin of https://hardware.research.yshicarrick.com/p/eagle4-arch-book/ . Labels: [M] measured, [D] derived, [P] projected.

## Track 1: FPGA drafter + GPU target

### C8 (0xE4B711C8) | 2026-09-24 | M01 · Baseline HEAD/TAIL split pair (C8+CF) | 153.4 tok/s [M] (pair batch-1)

- Bottleneck: the full EAGLE drafter layer did not fit or close timing on one V80.
- Added: a stage split. HEAD runs embedding, QKV, attention, O and MLP gate/up on a 128-lane 2-bit GEMV array; TAIL runs down-proj, fc_out, low-rank LM head, vocab tree and G8 rescore. The host relays the seam each level over PCIe Gen4 x8.
- Result: 153.4 tok/s [M F0001].
- role: HEAD
- timing: WNS +0.052 / WHS +0.011 ns
- state: tested
- result: 112.73 @ 2.22 w8; texts 50/50; G5 5.85x faster
- added / highlighted blocks: GEMV array (G4/G6/G7g/G7u); G5 attention + RoPE + softmax; G12 vocab sweep + TREE; Host relay CONS_HOST
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]

### CF (0xE4B711CF) | 2026-09-24 | M01 · Baseline HEAD/TAIL split pair (C8+CF) | 153.4 tok/s [M] (pair batch-1)

- Bottleneck: the full EAGLE drafter layer did not fit or close timing on one V80.
- Added: a stage split. HEAD runs embedding, QKV, attention, O and MLP gate/up on a 128-lane 2-bit GEMV array; TAIL runs down-proj, fc_out, low-rank LM head, vocab tree and G8 rescore. The host relays the seam each level over PCIe Gen4 x8.
- Result: 153.4 tok/s [M F0001].
- role: TAIL
- timing: WNS +0.015 / WHS +0.011 ns
- state: on card
- result: w2 153.4 @1.83 (best); w3 146.8 @1.97; w8 111 @2.21
- added / highlighted blocks: GEMV array (G4/G6/G7g/G7u); G5 attention + RoPE + softmax; G12 vocab sweep + TREE; Host relay CONS_HOST
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]

### E1 (0xE4B711E1) | 2026-09-24 | M02 · SEAM_WIDE burst export + KV-only attention skip (E1) | 178.16 tok/s [M] (pair batch-1)

- Bottleneck: a 32-b seam export and attention spent on KV-only rows.
- Added: SEAM_WIDE = 256-b AXI burst export (m_axi_out to HBM), and KVONLY_SKIP_ATTN.
- Result: 178.16 tok/s [M F0002]; export ~41 µs/item [D F0057].
- role: HEAD
- state: on card
- result: w2 178.2 (+17.6% vs C8), w3 175.9 @1.97 unprimed; primed w3 d2 118.96 @2.61 (priming −32%); unprimed w3 d3/5/8/15 = 2.08/2.14/2.15/2.15 flat; primed w3 d3/5/8 = 3.15/3.79/4.36 grow
- added / highlighted blocks: Seam export (SEAM_WIDE); G5 attention + RoPE + softmax
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]

### CD (0xE4B711CD) | 2026-09-24 | M03 · 2:1 fast array clock (CD) + tile overlap (I12)

- Bottleneck: the GEMV phases dominated the HEAD item at 66 MHz.
- Added: a synchronous 2:1 array clock, clk_usr_1 = 132 MHz, with tile overlap (ARRAY_FAST_TP).
- Result: 1.77× per item in xsim [M F0013], later corrected for double-counted tile overhead; CD routes +0.005 ns [M F0014].
- role: HEAD
- timing: WNS +0.005 / WHS +0.010 ns
- state: superseded (untested)
- result: superseded by E1 before flash
- added / highlighted blocks: 2:1 array clock 132 MHz; GEMV array (G4/G6/G7g/G7u)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]

### A9b (0xE4B711A9) | 2026-09-24 | M03 · 2:1 fast array clock (CD) + tile overlap (I12)

- Bottleneck: the GEMV phases dominated the HEAD item at 66 MHz.
- Added: a synchronous 2:1 array clock, clk_usr_1 = 132 MHz, with tile overlap (ARRAY_FAST_TP).
- Result: 1.77× per item in xsim [M F0013], later corrected for double-counted tile overhead; CD routes +0.005 ns [M F0014].
- role: HEAD
- timing: WNS 0.000 / WHS +0.002 ns
- state: tested
- result: [M] +3-6% vs A8; -8.6..-22% vs A5 (BR063) | same build id as its twin row but a separate build copy (own slot/md5) - not a duplicate row
- added / highlighted blocks: 2:1 array clock 132 MHz; GEMV array (G4/G6/G7g/G7u)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]

### E7 (0xE4B711E7) | 2026-09-26 | M04 · 256-lane HEAD GEMV array (E7 / L256) | 191.52 tok/s [M] (pair batch-1)

- Bottleneck: array throughput.
- Added: 256 lanes (from 128). Weights in URAM288 configured 32K×32, 4 URAM per lane = 1,024 banks; 256-b activation broadcast (16 × int16); 48-b scaled accumulators, 32-b lane outputs.
- Result: HEAD item 0.734× [D F0042]; pair 191.52 tok/s [M F0003].
- role: HEAD
- timing: WNS 0.000 / WHS +0.008 ns
- state: tested
- result: silicon 2026-09-26 with F3: texts mt w2 d2 50/50 = E1+CF, accept identical; HEAD G4/G6/G7 0.506x [M]; WNS 0.000 no hang (~40 runs); weight/scale drops 0
- added / highlighted blocks: GEMV array (G4/G6/G7g/G7u); URAM weight banks
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]

### F3 (0xE4B711F3) | 2026-09-26 | M04 · 256-lane HEAD GEMV array (E7 / L256) | 191.52 tok/s [M] (pair batch-1)

- Bottleneck: array throughput.
- Added: 256 lanes (from 128). Weights in URAM288 configured 32K×32, 4 URAM per lane = 1,024 banks; 256-b activation broadcast (16 × int16); 48-b scaled accumulators, 32-b lane outputs.
- Result: HEAD item 0.734× [D F0042]; pair 191.52 tok/s [M F0003].
- role: TAIL
- timing: WNS +0.280 / WHS +0.010 ns
- state: on card (tested)
- result: silicon 2026-09-26 with E7: GLM 546,000 -> 136,752 cyc (0.25x, ~814/child) and TREE 2,190,384 -> 1,350,384 (-38%) [M]; primed d3 TAIL busy 60.6 -> 41.7 ms
- added / highlighted blocks: GEMV array (G4/G6/G7g/G7u); URAM weight banks
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]

### A8 (0xE4B711A8) | 2026-09-26 | M05 · KV import (KVI) + host FASTGLUE | 234.62 tok/s [M] (pair batch-1)

- Bottleneck: unprimed levels; the card re-primed KV on every request.
- Added: KV import (KVI). The host writes verified KV rows to HBM and the HEAD loads them (S_KVL), plus host FASTGLUE.
- Result: 234.62 tok/s [M F0005].
- role: HEAD
- timing: WNS +0.002 / WHS +0.010 ns
- state: tested
- result: silicon: LANES 256, 0 drops, texts 50/50 = E1+CF; mt d2 w2 192.51 / w3 189.01 @1.98 (first arms); KVI L0 PASS | silicon best 234.62 tok/s with KVI + FASTGLUE=4 (main 2026-09-26) | 
- added / highlighted blocks: KVI KV import; Host KV writer (KVI)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]

### F4 (0xE4B711F4) | 2026-09-26 | M05 · KV import (KVI) + host FASTGLUE | 234.62 tok/s [M] (pair batch-1)

- Bottleneck: unprimed levels; the card re-primed KV on every request.
- Added: KV import (KVI). The host writes verified KV rows to HBM and the HEAD loads them (S_KVL), plus host FASTGLUE.
- Result: 234.62 tok/s [M F0005].
- role: TAIL
- timing: WNS +0.054 / WHS +0.010 ns
- state: on card (tested with A5)
- result: silicon: TREE -33.2% vs F3 (primed and unprimed), matches unit xsim; GLM = F3
- added / highlighted blocks: KVI KV import; Host KV writer (KVI)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]

### A5 (0xE4B711A5) | 2026-09-26 | M06 · Pipelined attention HEAD (A5 / ATTN_PIPE) | 268.01 tok/s [M] (pair batch-1)

- Bottleneck: serial attention on primed and KVI levels.
- Added: pipelined G5 attention (ATTN_PIPE, CSA par 6) on the 256-lane HEAD; 32 heads × 96, HEAD_PAR 8.
- Result: 268.01 tok/s [M F0006].
- role: HEAD
- timing: WNS 0.000 / WHS +0.001 ns
- state: tested
- result: [M] KVI+FG4 mt w2 d2 268.01 (1r), 264.8 (3r); fold v2.1 neutral d2, +2.5% d5, +1.4% d7 (BR057); protocol v1 37 cells
- added / highlighted blocks: G5 attention + RoPE + softmax
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]

### FA (0xE4B711FA) | 2026-09-27 | M07 · EARLY_A TAIL + EAGER/SPEC0 + KVI_FAN 312 MHz | 307.67 tok/s [M] (pair batch-1)

- Bottleneck: the chain latency TAIL → next HEAD item.
- Added: EARLY_A (the TAIL exports the next `a` vector under GLM..G8), eager survivor push, SPEC0, and KVI_FAN readers on a 312 MHz HBM clock (256-b AXI, 2 KB bursts, 8 outstanding).
- Result: 304.39 tok/s [M F0335]; 307.67 on a fast boot [M F0507].
- role: TAIL
- timing: WNS +0.069 / WHS +0.011 ns
- state: LANDED -> trantor pdi_inbox
- result: handoff /shared/scratch/chain_latency/handoff/; BR060; 18:48Z STEP5 PDI oneflip_p2s_tail_asgard_0xE4B711FA_66m_12f04f1b.pdi relayed, md5+magic verified on box1, NOTE_FA_20260927.md
- added / highlighted blocks: EARLY_A a-export; KVI_FAN readers 312 MHz; TREE_STOP / EAGER
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]

### AF (0xE4B711AF) | 2026-09-27 | M07 · EARLY_A TAIL + EAGER/SPEC0 + KVI_FAN 312 MHz | 307.67 tok/s [M] (pair batch-1)

- Bottleneck: the chain latency TAIL → next HEAD item.
- Added: EARLY_A (the TAIL exports the next `a` vector under GLM..G8), eager survivor push, SPEC0, and KVI_FAN readers on a 312 MHz HBM clock (256-b AXI, 2 KB bursts, 8 outstanding).
- Result: 304.39 tok/s [M F0335]; 307.67 on a fast boot [M F0507].
- role: HEAD
- timing: WNS -0.761 / WHS 0.000 ns
- state: LANDED olegvm3 11:19Z - TIMING FAIL (not relayed)
- result: staged md5-OK 19:1xZ asgard plain peak 72; expected FAIL enforce-build-manifest | moved asgard -> olegvm3 (vmpar2) 19:14Z for AG2 priority; withdrawn pre-admission 09:00Z: KVI_FAN 
- added / highlighted blocks: EARLY_A a-export; KVI_FAN readers 312 MHz; TREE_STOP / EAGER
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]

### A3 (0xE4B711A3) | 2026-09-27 | M08 · G4 embedding fold (EMBFOLD, 9 HBM masters)

- Bottleneck: G4 QKV, 3,570 cycles per HEAD item [M F0724].
- Added: embedding constant-fold (EMBFOLD) fetched through 9 parallel 256-b HBM masters.
- Result: fold A/B 264.3–264.8 [M]; dropped from the HM line, kept in the MP2/CL4 HEADs (G4_EMBFOLD = 1 [M strap]).
- role: HEAD
- timing: WNS 0.000 / WHS +0.010 ns
- state: built
- result: PDI 05:35Z olegvm2 ~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_asgard_0xE4B711A3_66m_1fd8740a.pdi; FAIL classes vs E1: +strap-watcher tooth (DEPTHN_GATE_WATCH.log) | 2026-09-26
- added / highlighted blocks: G4 EMBFOLD x9 masters
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]

### HM (0xE4B711BB) | 2026-09-27 | M08 · G4 embedding fold (EMBFOLD, 9 HBM masters)

- Bottleneck: G4 QKV, 3,570 cycles per HEAD item [M F0724].
- Added: embedding constant-fold (EMBFOLD) fetched through 9 parallel 256-b HBM masters.
- Result: fold A/B 264.3–264.8 [M]; dropped from the HM line, kept in the MP2/CL4 HEADs (G4_EMBFOLD = 1 [M strap]).
- role: HEAD
- timing: WNS -0.013 / WHS +0.002 ns
- state: BUILT - HOLD (timing FAIL) -> trantor pdi_inbox/HOLD_HM (DO NOT FLASH); HM-r2 re-route FAILED to close (WNS -0.038, 268 fail, no PDI, 06:25Z
- result: handoff /shared/scratch/max_head/handoff_hm (fork f320609c, MD5LIST 729d92c0); static: pyslang 0 err x6, pkg/BD dry-run OK; xsim HG0/HG2/HRED queued olegvm3; fit = AG2's (SLR1 slic
- added / highlighted blocks: G4 EMBFOLD x9 masters
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]

### HX1 (0xE4B711BD) | 2026-09-29 | M09 · Packed-DSP58 GEMV (HX1 DOT_E3) + FS/FSL TAIL | 323.3 tok/s [M] (pair batch-1)

- Bottleneck: LUT congestion in the HEAD's 2-bit CSA trees, and TAIL SLR1 flip-flop saturation.
- Added: DOT_E3. One DSP58 computes a·(w_L + 2²³·w_L+1) for a lane pair; 17 DSP58 per 32 products, latency 6, bit-exact. TAIL: serializer skip (FS) and R4 seams in LUTRAM (FSL).
- Result: 323.30 tok/s [M F0511], then 331.17 with TREE_STOP [M F0567].
- role: HEAD
- timing: WNS +0.004 / WHS +0.009 ns
- state: LANDED -> trantor pdi_inbox (QM relay 2026-09-28T09:35:11Z, md5+magic verified on box1) - FLASH CANDIDATE (user OK; pair FB2)
- result: handoff /shared/scratch/dsp_gemv/handoff_hx1 (HANDOFF_MD5 + fork MD5LIST OK asgard+olegvm3); queued 09:1xZ peak 72; race_class.sh arbitrates; STEP4c-e3 per-SLR block -> DSP-GEMV ag
- added / highlighted blocks: DOT_E3 packed DSP58; Seams r4/r5/r1/r8
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]

### FSL (0xE4B711CC) | 2026-09-29 | M09 · Packed-DSP58 GEMV (HX1 DOT_E3) + FS/FSL TAIL | 323.3 tok/s [M] (pair batch-1)

- Bottleneck: LUT congestion in the HEAD's 2-bit CSA trees, and TAIL SLR1 flip-flop saturation.
- Added: DOT_E3. One DSP58 computes a·(w_L + 2²³·w_L+1) for a lane pair; 17 DSP58 per 32 products, latency 6, bit-exact. TAIL: serializer skip (FS) and R4 seams in LUTRAM (FSL).
- Result: 323.30 tok/s [M F0511], then 331.17 with TREE_STOP [M F0567].
- role: TAIL
- timing: WNS +0.210 / WHS +0.010 ns
- state: RELAYED box1 pdi_inbox 23:34Z (auto v3)
- result: xsim [M]: G0 strap-0 vs FS 18/18 0 mism cycles identical; G1 strap-1 vs strap-0 18/18 0 mism cycles identical (SUM 477,390); RED idx^1 FIRED 2.33M mism first G7P; S0 OFF diag 1012 
- added / highlighted blocks: DOT_E3 packed DSP58; Seams r4/r5/r1/r8
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]

### HX1 BD (0xE4B711BD) | 2026-10-01 | M10 · LEVEL_PIPE HEAD (HX1 BD) + FG8/FGI TAIL + RELAY_LEAN | 337.22 tok/s [M] (pair batch-1)

- Bottleneck: serial export and the G8 tail on the per-level critical path.
- Added: LEVEL_PIPE (child k exports while child k+1 runs its front), FG8 (G8 rescore under the TREE sweep), ING_SILU_OVL, and host RELAY_LEAN.
- Result: HEAD item 184.0 µs b2b [M F0679]; FGI TAIL 11,211 cyc [M F0558]; pair 337.22 tok/s [M F0669].
- role: HEAD
- timing: WNS +0.004 / WHS +0.009 ns
- state: LANDED -> trantor pdi_inbox (QM relay 2026-09-28T09:35:11Z, md5+magic verified on box1) - FLASH CANDIDATE (user OK; pair FB2)
- result: handoff /shared/scratch/dsp_gemv/handoff_hx1 (HANDOFF_MD5 + fork MD5LIST OK asgard+olegvm3); queued 09:1xZ peak 72; race_class.sh arbitrates; STEP4c-e3 per-SLR block -> DSP-GEMV ag
- added / highlighted blocks: LEVEL_PIPE; FG8: G8 under TREE
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]

### FGI (0xE4B711AE) | 2026-10-01 | M10 · LEVEL_PIPE HEAD (HX1 BD) + FG8/FGI TAIL + RELAY_LEAN | 337.22 tok/s [M] (pair batch-1)

- Bottleneck: serial export and the G8 tail on the per-level critical path.
- Added: LEVEL_PIPE (child k exports while child k+1 runs its front), FG8 (G8 rescore under the TREE sweep), ING_SILU_OVL, and host RELAY_LEAN.
- Result: HEAD item 184.0 µs b2b [M F0679]; FGI TAIL 11,211 cyc [M F0558]; pair 337.22 tok/s [M F0669].
- role: TAIL
- timing: WNS 0.014 / WHS 0.010 ns
- state: RELAYED 2026-10-01 02:24Z (box1 verified); SUPERSEDED by FGIpx 746d2f94 (spare)
- result: [M xsim] -517 cyc/TAIL item (11728->11211), 0 mismatches; [P] -8.5 us/level, +0.3..0.4% batch 1; F0558-F0562 | [M] PDI: NEW 0 vs FG8Lax 26a5f241, FSLCHECK PASS, ISOCHECK PASS (iso_
- added / highlighted blocks: LEVEL_PIPE; FG8: G8 under TREE
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]

### TCLK75 (0xE4B71199) | 2026-10-01 | M11 · TAIL clock 66 -> 75 MHz (TCLK75) - RECORD | 339.12 tok/s [M] (pair batch-1)

- Bottleneck: the single TAIL kernel clock.
- Added: the same FGI RTL with the TAIL clock moved 66 → 75 MHz; routes WNS +0.016 / WHS +0.012 ns [M].
- Result: the RECORD, 339.12 tok/s [M F0686] vs GPU-only EAGLE 446–486 [M F0007/F0286].
- role: TAIL
- timing: WNS 0.016 / WHS 0.012 ns
- state: LANDED 2026-10-01 midyard (TCLK75md); PASS; relayed pdi_inbox
- result: [M] routed @75 MHz WNS +0.016/WHS +0.012; TCLKCHECK/FSLCHECK/ISOCHECK PASS; only new class OOC-WNS -0.031 (pre-route) vs FGI d197bac9
- added / highlighted blocks: TAIL clock 75 MHz
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]

### HQ_LV2 (0xE4B711D4) | 2026-10-01 | M12 · QSFP card link (Design D / CP1Q / LV2, link_dma v2)

- Bottleneck: host relay at 81–88 µs per hop [M].
- Added: a QSFP28 card link. Aurora 64B66B 4 × 25.78 Gb/s on GTM, AXIS 256 b at 402.8 MHz, and link_dma_eng v2 (4 BRAM TX buffers).
- Result: 39.9 µs per 70 KB hop [M xsim F0608]. Cables were not yet installed.
- role: HEAD
- timing: WNS 0.000 / WHS 0.000 ns
- state: LANDED 2026-10-01 20:37 EDT olegvm2; RELEASED 2026-10-01 ~22:00 EDT -> trantor FOUR_CARD/C_QSFP/LV2 (second QSFP option; host CP1Q_LINK_HOST
- result: [M] 0 real new FAIL vs HQ_DC_v4 (strap-watcher early-sample artifact: re-score finds g_depth_n/emb_gather present)
- added / highlighted blocks: QSFP28 Aurora 64B66B; link_dma_eng v2
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]

### TQ_LV2md (0xE4B711D5) | 2026-10-01 | M12 · QSFP card link (Design D / CP1Q / LV2, link_dma v2)

- Bottleneck: host relay at 81–88 µs per hop [M].
- Added: a QSFP28 card link. Aurora 64B66B 4 × 25.78 Gb/s on GTM, AXIS 256 b at 402.8 MHz, and link_dma_eng v2 (4 BRAM TX buffers).
- Result: 39.9 µs per 70 KB hop [M xsim F0608]. Cables were not yet installed.
- role: TAIL
- timing: WNS 0.004 / WHS 0.008 ns
- state: LANDED 2026-10-01 20:20 EDT midyard; RELEASED 2026-10-01 ~22:00 EDT -> trantor FOUR_CARD/C_QSFP/LV2 (second QSFP option; host CP1Q_LINK_HOST
- result: [M] 0 NEW / 0 DROPPED FAIL classes vs TQ_EB_v4 492a12fe; in-pipeline link G1..G7 INCONCLUSIVE as parent
- added / highlighted blocks: QSFP28 Aurora 64B66B; link_dma_eng v2
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]

### HQ_D2c (no build id) | 2026-10-03 | M13 · D2 card-side level loop | 348.9 tok/s [P] (pair batch-1)

- Bottleneck: host glue of ~226 µs per level.
- Added: the D2 card-side level loop. TAIL survivors and the parent `a` return over QSFP and the HEAD issues the next level without the host.
- Result: projected 348.9 tok/s [P F0644]; pair images and kit staged.
- added / highlighted blocks: D2 card-side level loop
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]

### TQ_D2 (0xE4B7119F) | 2026-10-03 | M13 · D2 card-side level loop | 348.9 tok/s [P] (pair batch-1)

- Bottleneck: host glue of ~226 µs per level.
- Added: the D2 card-side level loop. TAIL survivors and the parent `a` return over QSFP and the HEAD issues the next level without the host.
- Result: projected 348.9 tok/s [P F0644]; pair images and kit staged.
- role: TAIL
- timing: WNS 0.006 / WHS 0.003 ns
- state: LANDED 2026-10-02 02:16; PASS (gates+census) but NOT relayed: pair HEAD HQ_D2 FAIL
- result: [M] d2c strap1 G3/G4/G6/G6b PASS; D2/G7 zeros = regex artifact: census loop_seq_d2 + d2_eng/link_dma_eng_lp_v2/v2c + hop_card_rx present
- added / highlighted blocks: D2 card-side level loop
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]

### HQ_MP2s2 (0xE4B711EA) | 2026-10-04 | M14 · Two-link HEAD (MP2)

- Bottleneck: 4-card topologies need a second port on the HEAD.
- Added: a two-link HEAD (aur_link1 + link_dma1), with NoC SIs S20/S21 at 256 b.
- Result: HQ_MP2s2 routes 0.000 / +0.010 ns [M]; the pair is projected at 305–320 [P F0789].
- role: HEAD
- timing: WNS 0.000 / WHS 0.010 ns
- state: EARLY-LANDED 2026-10-04 03:05 from build dir; timing PASS (cascade hold did not recur); link gates pending official landing
- result: [M] 0 failing EP; 0 new FAIL vs HQ_MP2 (partial); BCAST body 0.000; qm18/early/gMP2s2
- added / highlighted blocks: 2nd Aurora link (MP2)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]

### TQ_LV2ae (0xE4B711D5) | 2026-10-04 | M14 · Two-link HEAD (MP2)

- Bottleneck: 4-card topologies need a second port on the HEAD.
- Added: a two-link HEAD (aur_link1 + link_dma1), with NoC SIs S20/S21 at 256 b.
- Result: HQ_MP2s2 routes 0.000 / +0.010 ns [M]; the pair is projected at 305–320 [P F0789].
- role: TAIL
- timing: WNS 0.019 / WHS 0.008 ns
- state: LANDED 2026-10-04 15:39 midyard; PASS (QM-20) - better-margin TAIL for the MP2 pair; relay to trantor FOUR_CARD/QSFP_MP armed (box1 unreacha
- result: [M] adjudicate_pdi.sh vs TQ_LV2md 96296e0d: 0 new / 0 dropped (18/18); g1fix3 link-gate replay (midyard mesh Midyard-844385, peak 33.6 GB): G1/G3/G4/G6/G6b/G7 PASS, G2/G5 waived ce
- added / highlighted blocks: 2nd Aurora link (MP2)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]

### SIBXE x2 (0xE4B71197) | 2026-10-05 | M15 · 4-card SIBX-E 2H2T over PCIe

- Bottleneck: batch-1 needs sibling lanes in parallel across cards; a deeper relay does not help.
- Added: SIBX-E 2H2T over PCIe, with two HEAD→TAIL chains and a host merge per level.
- SETBACK: on the SMI-slowed chassis, 4-card 50.00 vs 2-card 96.95 tok/s [M F0819]; KV load 9.08 ms. The record pair itself drops to 140.09 there [M F0764].
- role: TAIL
- timing: WNS 0.023 / WHS 0.007 ns
- state: LANDED 2026-10-01 15:59 EDT asgard; PASS -> relayed FOUR_CARD/B_SIBX (4-card only)
- result: [M] 0 real new FAIL vs SIBX ae08eb73: 1 new class 'STEP3: SIBX not wired into u_tree_level' = grep artifact (.SIBX_EN(SIBX), .SIBX_E_EN(SIBX_E)) u_tree_level at eagle4_drafter.sv:4
- added / highlighted blocks: SIBX-E sibling lanes; 2nd chain HEAD+TAIL (4-card)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]; SIBX-E sibling lanes [eager across 2 TAILs]; 2nd chain HEAD+TAIL (4-card) [2H2T / CL4]

### HQ_CL4 (0xE4B71200) | 2026-10-04 | M16 · CL4-L ring (HEAD<->HEAD QSFP merge) | 367 tok/s [P] (pair batch-1)

- Bottleneck: the host merge in SIBX-E.
- Added: the CL4-L ring. Two D2 chains merge on card over HEAD↔HEAD QSFP P2 (PART = 16 B + 8 B per entry, XITEM `a` = 6 KB).
- Result: HQ_CL4 +0.004 / TQ_CL4 0.000 ns WNS [M]; projected 366–368 tok/s [P]. Needs 3 cables.
- role: HEAD
- timing: WNS 0.004 / WHS 0.010 ns
- state: PDI written 2026-10-04 06:49 (tooth still running); PASS (HQ early adj) -> RELAYED as THE CL4 HEAD (NOTE_BOX1_CL4_HEAD_TO_USE.md, md5 verifi
- result: [M] F0796/F0797/F0802: 0 new FAIL vs HQ_MP2/HQ_D2_LV2/HQ_LV2 (gate table to PDI); census CL4/D2/G7 PASS; MP2 WHS-0.008 class did not recur
- added / highlighted blocks: CL4 HEAD fwd/merge; CL4 PART/XITEM; HEAD<->HEAD P2 merge (CL4)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]; SIBX-E sibling lanes [eager across 2 TAILs]; 2nd chain HEAD+TAIL (4-card) [2H2T / CL4]; CL4 HEAD fwd/merge [PART 16B+8B/entry]; CL4 PART/XITEM [msg 7/8]; HEAD<->HEAD P2 merge (CL4) [6 KB per level]

### TQ_CL4 (0xE4B71201) | 2026-10-04 | M16 · CL4-L ring (HEAD<->HEAD QSFP merge) | 367 tok/s [P] (pair batch-1)

- Bottleneck: the host merge in SIBX-E.
- Added: the CL4-L ring. Two D2 chains merge on card over HEAD↔HEAD QSFP P2 (PART = 16 B + 8 B per entry, XITEM `a` = 6 KB).
- Result: HQ_CL4 +0.004 / TQ_CL4 0.000 ns WNS [M]; projected 366–368 tok/s [P]. Needs 3 cables.
- role: TAIL
- timing: WNS 0.0 / WHS 0.005 ns
- state: LANDED 2026-10-04 06:36:00 EDT olegvm2; HOLD (not relayed)
- result: [M] vs TQ_D2_LV2 0xE4B7119F e759f0a6 (class new only if new vs ALL): NEW 5 (real 5); adjudication /shared/scratch/cl4/hq_cl4h/watch/TQ_CL4_oneflip_p2s_tail_asgard_0xE4B71201_66m_e4
- added / highlighted blocks: CL4 HEAD fwd/merge; CL4 PART/XITEM; HEAD<->HEAD P2 merge (CL4)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]; SIBX-E sibling lanes [eager across 2 TAILs]; 2nd chain HEAD+TAIL (4-card) [2H2T / CL4]; CL4 HEAD fwd/merge [PART 16B+8B/entry]; CL4 PART/XITEM [msg 7/8]; HEAD<->HEAD P2 merge (CL4) [6 KB per level]

### SG_HEAD (0xE4B71206) | 2026-10-05 | M17 · Runtime shape table (SHAPE_RT)

- Bottleneck: the drafter shape is fixed at compile time.
- Added: SHAPE_RT, a runtime GEMV shape table at CSR 0x200/0x204 (10-b AXI-Lite).
- Result: 3B is cycle-exact, 0 of 1,292,916 cycles differ [M]. Path to 8B (~303) and Qwen-27B (~172 tok/s) on 4-card TP [P].
- role: HEAD
- state: BUILDING 2026-10-05 (mesh pdi_build_head_sg peak 72; drain-wait behind HX3TF at submit) -- gate FH_SG4b vs FH_REF3 cycle-exact PASS [M]
- added / highlighted blocks: SHAPE_RT table; SHAPE_RT (TAIL)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]; SIBX-E sibling lanes [eager across 2 TAILs]; 2nd chain HEAD+TAIL (4-card) [2H2T / CL4]; CL4 HEAD fwd/merge [PART 16B+8B/entry]; CL4 PART/XITEM [msg 7/8]; HEAD<->HEAD P2 merge (CL4) [6 KB per level]; SHAPE_RT table [0x200/0x204]; SHAPE_RT (TAIL) [0x200/0x204]

### SG_TAIL (0xE4B71207) | 2026-10-05 | M17 · Runtime shape table (SHAPE_RT)

- Bottleneck: the drafter shape is fixed at compile time.
- Added: SHAPE_RT, a runtime GEMV shape table at CSR 0x200/0x204 (10-b AXI-Lite).
- Result: 3B is cycle-exact, 0 of 1,292,916 cycles differ [M]. Path to 8B (~303) and Qwen-27B (~172 tok/s) on 4-card TP [P].
- role: TAIL
- timing: WNS 0.015 / WHS 0.008 ns
- state: LANDED 2026-10-05 18:15 EDT olegvm3; PASS (QM-22); pair relay to SHAPE_GENERIC/PAIR2 waits for SG_HEAD
- result: [M] 0 real new FAIL vs TCLK75 df7af3df (strap-watcher race, enorm watcher INCONCLUSIVE->PASS); qm22/SG_PAIR2/SG_TAIL_VERDICT.txt
- added / highlighted blocks: SHAPE_RT table; SHAPE_RT (TAIL)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]; SIBX-E sibling lanes [eager across 2 TAILs]; 2nd chain HEAD+TAIL (4-card) [2H2T / CL4]; CL4 HEAD fwd/merge [PART 16B+8B/entry]; CL4 PART/XITEM [msg 7/8]; HEAD<->HEAD P2 merge (CL4) [6 KB per level]; SHAPE_RT table [0x200/0x204]; SHAPE_RT (TAIL) [0x200/0x204]

### SG_CL4H (0xE4B71208) | 2026-10-05 | M17 · Runtime shape table (SHAPE_RT)

- Bottleneck: the drafter shape is fixed at compile time.
- Added: SHAPE_RT, a runtime GEMV shape table at CSR 0x200/0x204 (10-b AXI-Lite).
- Result: 3B is cycle-exact, 0 of 1,292,916 cycles differ [M]. Path to 8B (~303) and Qwen-27B (~172 tok/s) on 4-card TP [P].
- role: HEAD
- timing: WNS 0.000 / WHS 0.000 ns
- state: LANDED 2026-10-05 19:20 EDT (verdicts 20:50); PASS at zero margin post_route_phys_opt WNS 0.000 / WHS 0.000, 0 failing EP (route_design was 
- result: [M] 0 real new FAIL vs HQ_CL4 5a4f219d (strap-watcher race, async EVAC timeout); qm22/SG_CL4H/SG_CL4H_v1_VERDICT.txt
- added / highlighted blocks: SHAPE_RT table; SHAPE_RT (TAIL)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]; SIBX-E sibling lanes [eager across 2 TAILs]; 2nd chain HEAD+TAIL (4-card) [2H2T / CL4]; CL4 HEAD fwd/merge [PART 16B+8B/entry]; CL4 PART/XITEM [msg 7/8]; HEAD<->HEAD P2 merge (CL4) [6 KB per level]; SHAPE_RT table [0x200/0x204]; SHAPE_RT (TAIL) [0x200/0x204]

### SG_CL4T (0xE4B71209) | 2026-10-05 | M17 · Runtime shape table (SHAPE_RT)

- Bottleneck: the drafter shape is fixed at compile time.
- Added: SHAPE_RT, a runtime GEMV shape table at CSR 0x200/0x204 (10-b AXI-Lite).
- Result: 3B is cycle-exact, 0 of 1,292,916 cycles differ [M]. Path to 8B (~303) and Qwen-27B (~172 tok/s) on 4-card TP [P].
- role: TAIL
- state: BUILDING 2026-10-05 16:41 EDT mesh olegvm2-1424345 peak 100 (QM-22 submit_when_free after TQ_CL4s2 exit) -- gate COMPARE_TGL_CL4_levers/v2 P
- added / highlighted blocks: SHAPE_RT table; SHAPE_RT (TAIL)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KVI KV import [KV_LOAD from HBM]; EARLY_A a-export [6 KB under GLM..G8]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; TREE_STOP / EAGER [early level exit]; G4 EMBFOLD x9 masters [9x256b HBM]; DOT_E3 packed DSP58 [17 DSP/32 prod, LAT6]; LEVEL_PIPE [export || next child]; FG8: G8 under TREE [overlap]; TAIL clock 75 MHz [usr_0 66->75]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 2nd Aurora link (MP2) [link_dma1, P2]; SIBX-E sibling lanes [eager across 2 TAILs]; 2nd chain HEAD+TAIL (4-card) [2H2T / CL4]; CL4 HEAD fwd/merge [PART 16B+8B/entry]; CL4 PART/XITEM [msg 7/8]; HEAD<->HEAD P2 merge (CL4) [6 KB per level]; SHAPE_RT table [0x200/0x204]; SHAPE_RT (TAIL) [0x200/0x204]

### KVONC ATT (0x130C) (0xE4B7130C) | 2026-11-08 | 4C1 · KV_ONCARD ATT (with MLPV 0x128B r2) | 810.1 tok/s [M] (Llama-3.2-3B humaneval, 4 cards (talon2), KVONC off)

- Bottleneck: host-side KV commit ops on every step.
- Added: KV_ONCARD strap 1 on the 0x129C dual-model ATT, so fused KV-only items commit KV on the card (CTX_WIN 768).
- Result: Llama 810.10 tok/s @ 5.88 and Qwen3.8-27B 248.19 tok/s @ 7.14 [M]. KVONC host fix A/B: 749.57 vs 738.32 tok/s @ 5.49 (+1.52 %) [M]; both below 810.10 because the KV-commit stack runs at accept 5.49.
- Status: tested on silicon (talon2).
- role: ATT pair (talon2 17/3d) + MLPV 0x128B r2 (50/63)
- timing: WNS +0.040 / WHS +0.012 ns
- state: tested 2026-11-08; flashed 17:44Z
- result: Llama 810.10 @ 5.88; Qwen 248.19 @ 7.14; KVONC K 749.57 vs B 738.32 @ 5.49; texts K=B 80/80
- pdi md5: 02b1efd5
- source: FOURCARD_DESIGN_20261016/4C1_DELIVERY_TRACKER.md:1372-1373,1414; KVONC_HOSTFIX_20261108.md §5; ADJUDICATION_0x130C_20261108.md
- added / highlighted blocks: KV_ONCARD (KVONC) on ATT
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; G4 EMBFOLD x9 masters [9x256b HBM]; KV_ONCARD (KVONC) on ATT [fused KV-only items, CTX_WIN 768]

### KV_BANKS ATT (0x1334) (0xE4B71334) | 2026-11-08 | 4C1 · KV_BANKS two-stream ATT

- Bottleneck: one KV context per ATT card, so a second stream replays its KV at every switch (N2 measured 239.67 tok/s aggregate).
- Added: KV_BANKS 2 x 512 slots per head with a scoped 0x154 clear, G6_HOLD_GATE (holds raw-tile issue during a link stall) and the K8W1 floorplan. No KV_ONCARD.
- Result: xsim G6 hold loses 0/36 tiles at stalls of 100 and 512 cycles, +0 cycles; plain SV reference 0 differing words [M].
- Status: flash-ready (adjudicated, 0 flash-relevant FAILs); not yet run on silicon.
- role: ATT (dual model), pairs with MLPV 0x128B r2
- timing: WNS +0.041 / WHS +0.011 ns
- state: flash-ready 2026-11-08
- pdi md5: dd975b24
- note: run with EAGLE4_CTX_WIN=512
- source: FOURCARD_DESIGN_20261016/ADJUDICATION_0x1334_20261108.md; BUILD_VERIFY_20261109/V_0x1334_0x1322.md
- added / highlighted blocks: KV_BANKS 2x512 + G6 hold
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KV_BANKS 2x512 + G6 hold [2 resident streams, 0x154 scoped clear]

### RET_SER R=2 ATT (0x1322) (0xE4B71322) | 2026-11-08 | 4C1 · RET_SER return serializer on KV_BANKS | 759.02 tok/s [M] (Llama-3.2-3B humaneval 80x256, 4 cards (talon2), CTX_WIN 512, mean of 3)

- Bottleneck: SLR1-SLR0 wires. The full-width result return of 0x1309 needed 18,767 of 18,870 SLLs.
- Added: RET_SER R=2, which serializes the array result return, on the K8W1 KV_BANKS ATT.
- Result: full-drafter replay bit-identical 71/71 on both streams; cost +10,736 busy cycles (+1.04 %), +2,684 on each of G4/G6/G7g/G7u [M xsim].
- Status: tested on silicon (talon2, 2026-11-09): Llama 759.02 @ 5.45, Qwen 233.73 @ 6.72 [M].
- role: ATT (dual model) + MLPV 0x128B r2
- timing: WNS +0.042 / WHS +0.012 ns, 0 route errors
- result: Llama 759.02 @ 5.45; Qwen 233.73 @ 6.72
- correctness: Llama 13/13, Qwen 9/9 trees = 0x1339; texts 80/80 = 0x1334
- pdi md5: 45897272
- source: FOURCARD_DESIGN_20261016/SILICON_0x1322_20261109.md §0; ADJUDICATION_0x1322_20261108.md
- added / highlighted blocks: RET_SER result return
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KV_BANKS 2x512 + G6 hold [2 resident streams, 0x154 scoped clear]; RET_SER result return [serialized array return, R=2]

### R3_EARLY ATT (0x1339) (0xE4B71339) | 2026-11-08 | 4C1 · R3_EARLY + registered 0x90 fold | 764.64 tok/s [M] (Llama-3.2-3B humaneval 80x256, 4 cards (talon2), CTX_WIN 512)

- Bottleneck: R3 waited for G6 done; the 0x90 debug XOR fold was a 135-LUT6 chain (F0 post-place -17.376).
- Added: on 0x1334, R3_EARLY 1 (R3 launches at G6 start) and TIMING_HARD class A (registered 0x90 fold).
- Result: identity replay vs 0x1334 PASS 38/38 and 33/33, X count 0, R3 RED bites; plain SV reference 0 differing words [M]. Post-place WNS +0.002 (route not final). Silicon 2026-11-09: Llama 764.64 @ 5.45 and Qwen 234.94 @ 6.72; goldens and texts = 0x1334 [M].
- Status: tested on silicon (talon2, 2026-11-09); the 11-08 F0b failure did not reproduce on this image.
- role: ATT (dual model) + MLPV 0x128B r2
- result: Llama 764.64 @ 5.45; Qwen 234.94 @ 6.72 (CTX_WIN 512)
- correctness: goldens Llama 8/8, Qwen 9/9 = 0x1334; texts 80/80 = 0x1334
- state: tested 2026-11-09
- source: FOURCARD_DESIGN_20261016/SILICON_0x1339_20261109.md §0; R3EARLY_0x1339_20261108.md §5
- added / highlighted blocks: R3_EARLY + registered 0x90 fold
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; 2:1 array clock 132 MHz [usr_1 132 MHz]; KV_BANKS 2x512 + G6 hold [2 resident streams, 0x154 scoped clear]; R3_EARLY + registered 0x90 fold [R3 launch at G6 start; TIMING_HARD A]

### Q5120 ATT L256 (0x1344) (0xE4B71344) | 2026-11-08 | 4C1 · Qwen3.8-27B drafter Q5120 ATT (TP2)

- Bottleneck: the Qwen3.8-27B drafter (H 5120, 40 heads, inter 16384) does not fit the v1 split: 1,536 array URAM against 1,248 in SLR1+SLR2.
- Added: v0 split (ATT = G4, G5, G6; G7 moves to MLPV), 20x128 heads per rank, LUT-GEMM A8 at L256, K_MAX 10240, SCALE_FRAC 16, RET_SER.
- Result: full-drafter xsim root + L1 bit-exact vs the bit-true golden; ATT item 17,773 cycles = 269 µs at 66 MHz [M xsim].
- Status: landed 2026-11-09, quarantined: post-route WNS -2.616 / WHS +0.000; constraint retime with F1+F2 -0.545 / +0.001 [M]. Closure image 0x1347.
- role: ATT TP2
- timing: post-route WNS -2.616 / WHS +0.000; retime F1+F2 WNS -0.545 / WHS +0.001 ns
- state: quarantined (negative WNS)
- xsim: G4/KV/G5/G6sum/h1 0 differing, both ranks
- source: QWEN27B_DRAFTER_20261020/TIMING_0x1344_20261109.md:107-109; 4C1_DELIVERY_TRACKER.md:1457
- added / highlighted blocks: Q5120 ATT geometry (LUTG A8)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; Q5120 ATT geometry (LUTG A8) [H 5120, 20x128/rank, K_MAX 10240]; KV_BANKS 2x512 + G6 hold [2 resident streams, 0x154 scoped clear]; RET_SER result return [serialized array return, R=2]

### Q5120 ATT L512 (0x1345) (0xE4B71345) | 2026-11-08 | 4C1 · Qwen3.8-27B drafter Q5120 ATT (TP2)

- Bottleneck: the L256 ATT item is 17,773 cycles, of which G4 QKV is 10,712.
- Added: LANES 512 on the same Q5120 ATT source (att_g6_tp skew 6-bit).
- Result: no card-level xsim at L512; evidence is the L256 xsim and the stage gate.
- Status: staged, not launched.
- role: ATT TP2
- state: staged (buildx04 / buildx02)
- source: QWEN27B_DRAFTER_20261020/QWEN5120_P1_20261108.md §4; BUILD_VERIFY_20261109/V_Q5120_ATT.md finding 3
- added / highlighted blocks: Q5120 ATT geometry (LUTG A8)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; Q5120 ATT geometry (LUTG A8) [H 5120, 20x128/rank, K_MAX 10240]; KV_BANKS 2x512 + G6 hold [2 resident streams, 0x154 scoped clear]; RET_SER result return [serialized array return, R=2]

### Q5120 TP4 ATT (0x1346) (0xE4B71346) | 2026-11-09 | 4C1 · Qwen3.8-27B drafter Q5120 ATT (TP4, 7-card)

- Bottleneck: the TP2 ATT carries all of QKV at K 10240 per card.
- Added: a 7-card drafter rank: 4 ATT (TP4) + 3 MLPV, 10x128 heads per card, G4 N 3840, G6 K 1280, hypercube G6 sum, one-die array.
- Result: OOC array alone routed WNS +0.294 / WHS +0.010 [M]; 4-rank xsim h1 0/5,120 differing on all 4 ranks [M].
- Status: landed 2026-11-09, quarantined: post-route WNS -2.565 / WHS +0.006 [M]; superseded by 0x134A (TP4 closure) and 0x1347 (TP2 closure).
- role: ATT TP4 rank
- timing: post-route WNS -2.565 / WHS +0.006 ns
- state: quarantined; superseded by 0x134A / 0x1347
- source: 4C1_DELIVERY_TRACKER.md:1466,1461; QWEN27B_DRAFTER_20261020/QWEN5120_TP4_ATT_20261108.md
- added / highlighted blocks: TP4 ATT rank (7-card)
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; Q5120 ATT geometry (LUTG A8) [H 5120, 20x128/rank, K_MAX 10240]; TP4 ATT rank (7-card) [4 ATT + 3 MLPV; 10x128/card, hypercube G6 sum]

### Q5120 MLPV L512 (0x1348) (0xE4B71348) | 2026-11-09 | 4C1 · Qwen3.8-27B drafter Q5120 MLPV

- Bottleneck: the MLPV item at Q5120 (G7g/G7u, G9, 248,320-id vocab) and the 788 MB-class fp16 lm-head gather.
- Added: on the 0x1342 MLPV source, LUT-GEMM A8 at L512, TREE_P 64, vocab shard per rank, and G8 exact rescore on 2-bit lm_head rows from HBM (1,408 B/row); fp16 G8 dropped.
- Result: unit gates 0 differing (46,080 / 46,080 / 90,320); full-MLPV pair xsim not yet green at verify time.
- Status: staged (buildx04); flash quarantined until the pair xsim and post-synth checks pass.
- role: MLPV
- state: staged
- open: full 349.6 MB lm-head table per card; BD-wrapper grep
- source: QWEN27B_DRAFTER_20261020/QWEN5120_P2_MLPV_20261108.md; BUILD_VERIFY_20261109/V_Q5120_MLPV.md
- added / highlighted blocks: G8 rescore on 2-bit lm_head rows
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; Q5120 ATT geometry (LUTG A8) [H 5120, 20x128/rank, K_MAX 10240]; G8 rescore on 2-bit lm_head rows [HBM 1,408 B/row, TREE_V_MAX 248320]

### Q5120 MLPV L256 (0x1349) (0xE4B71349) | 2026-11-09 | 4C1 · Qwen3.8-27B drafter Q5120 MLPV

- Bottleneck: as 0x1348; this is the L256 / TREE_P 32 arm.
- Added: Q5120 MLPV at L256, TREE_P 32, same 2-bit-row G8 rescore and vocab shard.
- Result: full-MLPV pair xsim root green on both ranks, 381,760 words 0 differing; planted id 140001 green, rank-bit RED fails [M].
- Status: building (buildx00, launched 2026-11-09 00:49Z); quarantined until md>=2 + REDs + post-synth grep.
- role: MLPV
- state: building
- source: BUILD_VERIFY_20261109/V_Q5120_MLPV.md; QWEN5120_P2_MLPV_20261108.md; 4C1_DELIVERY_TRACKER.md:1425
- added / highlighted blocks: G8 rescore on 2-bit lm_head rows
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; Q5120 ATT geometry (LUTG A8) [H 5120, 20x128/rank, K_MAX 10240]; G8 rescore on 2-bit lm_head rows [HBM 1,408 B/row, TREE_V_MAX 248320]

### BATCH + prefill ATT (merged) (no build id) | 2026-11-09 | 4C1 · BATCH_MODE + prefill pair (design)

- Bottleneck: one stream per card; switching streams re-stages KV (N2 239.67 tok/s aggregate measured, 78.9 % of card time re-stage).
- Added (design): runtime BATCH_MODE at 0x240 (0 single, 1 max-batch tree-grain, 2 balanced item-grain) over ONE KV bank/position map shared with PF_V prefill; base pf_kvprefill_20261109 (D9 LUTG + SIBW + PF_V).
- Result: none on silicon. MODE 0 must stay bit- and cycle-exact to the parent.
- Status: design; ATT built by box1-qwen-junior.
- role: ATT
- state: design
- source: BATCH_BUBBLES_20261109/IMAGE_A_MAXBATCH_NOLOSS.md §0; BATCH_BUBBLES_20261109/00_SYNTHESIS.md; 4C1_DELIVERY_TRACKER.md:1422
- added / highlighted blocks: BATCH_MODE 0x240 KV bank map; PF_V on-card KV prefill
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; BATCH_MODE 0x240 KV bank map [modes 0/1/2; 1x1024 .. 8x256 slots]; PF_V on-card KV prefill [4th arm of the one KV write port]

### BATCH MLPV (merged) (no build id) | 2026-11-09 | 4C1 · BATCH_MODE + prefill pair (design)

- Bottleneck: the MLPV is tree-serial (one frontier per card) and its S1 overlap mode is locked off.
- Added (design): stream id in the hop frame header [29:28], per-stream OUT-DMA base + stride, tree-id tag and S1 overlap, all runtime-gated.
- Result: none on silicon.
- Status: design; MLPV built by box1-two-v2.
- role: MLPV
- state: design
- source: BATCH_BUBBLES_20261109/IMAGE_A_MAXBATCH_NOLOSS.md §0.3; 00_SYNTHESIS.md; 4C1_DELIVERY_TRACKER.md:1422
- added / highlighted blocks: MLPV stream tag + S1 overlap
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; BATCH_MODE 0x240 KV bank map [modes 0/1/2; 1x1024 .. 8x256 slots]; PF_V on-card KV prefill [4th arm of the one KV write port]; MLPV stream tag + S1 overlap [per-stream OUT base, runtime-gated]

### F3t G5_OVL ATT (0x13F9) (0xE4B713F9) | 2026-11-09 | 4C1 · E2b + G5_OVL 1 + G6_REPLAY_DB 1 + TIMING_HARD 2'b11 | 831.14 tok/s [M] (Llama-3.2-3B humaneval 80x256, 4 cards (talon2), w2 d7, K_C 64; track-1 record)

- Bottleneck: G5 runs its two head groups (16 heads / HEAD_PAR 8) one after the other.
- Added: G5_OVL (overlaps the two head-group passes), G6_REPLAY_DB and TIMING_HARD 2'b11 on E2b.
- Result: Llama 831.14 / 830.09 @ 5.89 vs same-day E2b mean 814.44 (+1.99 %); Qwen 249.62 / 249.04 @ 7.14 (inside E2b noise). G5 per tree -30.8 %, tree issue->done p50 -8.3 % [M]. Texts 80/80 and goldens identical to E2b.
- Status: tested on silicon (talon2); track-1 record.
- role: ATT pair (talon2 17/3d) + MLPV 0x1342
- timing: FINAL_TRUE WNS +0.040 / WHS +0.011 ns
- result: Llama 831.14 @ 5.89; Qwen 249.62 @ 7.14
- card census: G5 751.55 -> 519.98 us/tree (-30.8 %); issue->done p50 2,422 -> 2,220 us (-8.3 %), Llama
- pdi md5: 97d8445c
- source: TRACK1_RECORD_F3T_20261109/EVALUATION.md; DRAFTER_SPEED_RESEARCH_20261107/F3T_FLASH_ADJUDICATION_20261109.md
- added / highlighted blocks: G5_OVL head-group overlap + G6_REPLAY_DB
- blocks present: H200 target verify (SGLang) [PCIe Gen4 x16]; Host relay CONS_HOST [Python/C relay]; Host KV writer (KVI) [KV rows -> HBM]; PCIe Gen4 x8 QDMA (HEAD); PCIe Gen4 x8 QDMA (TAIL); AXI-Lite CSR [32b data, 9b word addr]; Top FSM [S_G1..S_LP_BAR]; G1 embedding gather [m_axi_emb 256b]; RMSNorm [256b chunk, H=3072]; GEMV array (G4/G6/G7g/G7u) [LANES 256 x K_PAR16 x int16; acc 48b]; URAM weight banks [32Kx32, 4 URAM/lane (1024)]; Scale BRAM Q3.13 [16b, group 128]; G5 attention + RoPE + softmax [32x96, HEAD_PAR 8]; KV cache [K URAM (176) / V BRAM, CTX 768]; KVI KV import [KV_LOAD from HBM]; KVI_FAN readers 312 MHz [usr_2, 256b, burst 2KB]; Seam export (SEAM_WIDE) [m_axi_out 256b]; HBM (HEAD) [16 ch, NoC 256b SIs]; Ingest bridge [2240 x 256b beats]; Seams r4/r5/r1/r8 [FF -> LUTRAM (R4)]; SiLU x up (G7p) [8192, Q6.12]; G9 down + R7 (G9X/G9Y) [LANES 128, 2-bit]; G11 fc_out + RMSNorm [a: 192x256b Q6.9]; GLM low-rank LM-down [3072->256, 64 lanes]; G12 vocab sweep + TREE [1002 tiles, TREE_P 32]; G8 exact fp16 rescore [G8_NCH 4 x 256b, K_C 64]; Level FSM top-W [W x 8 children]; URAM weight banks [640 + 256 g12x9m]; HBM (TAIL) [LM-head fp16 rows, seams]; QSFP28 Aurora 64B66B [4 x 25.78 Gb/s, AXIS 256b @402.8 MHz]; link_dma_eng v2 [4 TX bufs, MAXB 64, 39.9 us/70 KB]; D2 card-side level loop [loop_seq_d2]; 4C1 split: ATT pair + MLPV pair [ATT TP2 heads/rank; MLPV vocab shard; QSFP hop]; D9 LUT-GEMM A8 array (ATT) [multiply-free 2-bit lanes, LANES 512 or 256, K_PAR 16, A8Q]; RET_SER result return [serialized array return, R=2]; G5_OVL head-group overlap + G6_REPLAY_DB [16 heads / HEAD_PAR 8 = 2 groups, passes overlapped]

## Track 2: FPGA target + drafter, HBM streaming

### SLR1233np32s6_200 (0xE4B71233) | 2026-10-09 | HBMS NP32 base

- Three-SLR rtl_s6 floorplan: ports 0-15 far tile (SLR2, local BLI, KV ports), ports 16-31 near tile (SLR0).
- NP32 at 200 MHz; the first image of the HBM-streaming engine.
- Result: PASS QM-34, relayed do-not-flash [M F1174].
- state: PASS QM-34, relayed do-not-flash
- clock: hbms_clk 200 MHz
- added / highlighted blocks: HBM port engines x32; HBM2e 32 GB
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]

### SLR1235np32s6_225 (0xE4B71235) | 2026-10-10 | NP32 @ 225 MHz

- Same RTL as 0x1233 (13 of 14 sources md5-identical); clock raised to 225 MHz.
- Margin upgrade s6u: true hbms_clk +0.104 ns [D] vs +0.040.
- Result: PASS QM-35 / QM-36 [M F1194, F1200].
- timing: WNS +0.040 / WHS +0.010 ns [M]
- clock: hbms_clk 225 MHz
- added / highlighted blocks: Clocks
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]

### R8C4 0x1250 b1 (0xE4B71250) | 2026-10-11 | RING8 C4 one-image

- 0x1235 + crumbc C4, flashed on all 8 talon cards (partition 0).
- SETBACK: ATTN_MODE0 / layer gates fail on silicon (a0 lane-mapping hazard in hbml_attn) [M].
- No transceivers: ring ports are an internal loopback, so the ring cannot carry traffic.
- state: PASS QM-38; on all 8 talon cards, partition 0
- silicon: a0/layer FAIL [M]
- added / highlighted blocks: internal loopback; Attention
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; internal loopback [0x1250: ring ports end in a loopback FIFO; no transceivers]

### R8_C4_attnfix s2 (0x1254) (0xE4B71254) | 2026-10-15 | attnfix, serving image | 28.0 tok/s [M] (Llama-3.1-8B, one V80)

- hbml_attn lane fix + Explore/AggressiveExplore route: WNS +0.002 / WHS +0.010 ns [M].
- Ladder 52/52 on all 8 talon cards [M F1257]. Serves real Llama models with no new PDI.
- Measured batch-1: 70B on 8 cards 2.34 · 8B 28.0 · 3B 54.8 · 1B 155.6 · TriLM-99M 1,018 tok/s [M].
- PDI: 259a2800
- silicon: 52/52 x 8 cards [M]
- 70B, 8 cards PP: 2.34 tok/s [M]
- 3B x 8 streams: 439 tok/s batch-8 aggregate [M]
- model switch: 3.1-3.9 ms [M]
- added / highlighted blocks: Attention; GEMV engine; 8 cards: pipeline-parallel via host
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; 8 cards: pipeline-parallel via host [70B: 10 layers / card, residual over MMIO]

### ring-fix link image (0x12B7) (0xE4B712B7) | 2026-10-14 | ring only

- Link-only image: Aurora ring on cages 0/3 with RX pacing + heal (hop_aurora_cdc RX FIFO overflow fixed).
- BIST PASS 16/16 at len 64 / 256 / 960; raw 99.84 Gb/s; BER <= 5e-14 [M].
- RING_MAP OK; ring order 17 > 3d > 50 > 63 > cf > e1 > 99 > bd.
- BIST: 16/16 PASS [M]
- raw: 99.84 Gb/s [M]
- added / highlighted blocks: QSFP28 Aurora ring core (cages 0 / 3); 8-card ring 17>3d>50>63>cf>e1>99>bd
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; QSFP28 Aurora ring core (cages 0 / 3) [4 x 25.78 Gb/s per direction, pace 8 + heal, DROP_RST; 99.84 Gb/s raw [M]]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]

### R8_HBMSRING_merge (0x1255) (0xE4B71255) | 2026-10-15 | pathfinder, building

- 0x1254 s2 HBMS + the talon ring core in one image, side by side.
- RTL check: HBMS ring ports end in hbml_core's loopback FIFO, so there is no data path to the ring [M, RTL].
- Worth: floorplan/timing pathfinder + ladder + ring BIST in one image. 4 seeds building.
- state: BUILDING: 4 seeds (olegvm2, olegvm3, midyard)
- risk: SLR1 slices ~98-99%
- added / highlighted blocks: QSFP28 Aurora ring core (cages 0 / 3); no data path between HBMS and the ring in this image
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; QSFP28 Aurora ring core (cages 0 / 3) [4 x 25.78 Gb/s per direction, pace 8 + heal, DROP_RST; 99.84 Gb/s raw [M]]; no data path between HBMS and the ring in this image [pathfinder: floorplan + timing + ladder + BIST only]

### R8_HBMSRING_AR8 (0x1256) (0xE4B71256) | 2026-10-15 | design: true 8-card TP8 | 206 tok/s [P] (Llama-3.2-3B TP8, 8 cards)

- Adds the HBMS-AR <-> ring bridge and widens ALLRED to P=8 (cfg_p 3 -> 4 bits).
- 3B TP8 per-card programs + loader are ready and bit-exact in emulation; 56 all-reduces per token.
- Projected ~206 tok/s batch-1 at 12.6 us per all-reduce [P].
- state: DESIGN (RTL in progress)
- runtime: tp_serve.py ready, emulator bit-exact [M emu]
- added / highlighted blocks: HBMS-AR <-> ring bridge; 8-card ring 17>3d>50>63>cf>e1>99>bd
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; QSFP28 Aurora ring core (cages 0 / 3) [4 x 25.78 Gb/s per direction, pace 8 + heal, DROP_RST; 99.84 Gb/s raw [M]]; HBMS-AR <-> ring bridge [async FIFO 225 -> 150 MHz, <=256-beat frames, ALLRED P=8]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]

### R8_C4_np64w8 (0x1257) (0xE4B71257) | 2026-10-15 | design: 64 ports | 97 tok/s [P] (Llama-3.2-3B, one V80)

- Decode is >= 92% W8 weight streaming, capped by the 32-port beat rate (230 GB/s) [D on M profile].
- 64 ports (W8-only if area requires), same ISA, margin RTL of 0x125C.
- Projected 3B on one card ~95-100 tok/s [P].
- state: DESIGN: feasibility first
- bottleneck: port beat rate, HBM stalls 2-4% [M]
- added / highlighted blocks: HBM port engines x64; HBM2e 32 GB
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; HBM port engines x64 [NP64 W8, ~2x beat rate [P]]

### RING8 v8 (0x1262) (0xE4B71262) | 2026-10-17 | v8: TP8 ring all-reduce, serving image | 169.1 tok/s [M] (Llama-3.2-3B TP8, 8 cards (VPU4))

- 0x1256-family recovery: HBMS-AR <-> ring bridge on the 8-ring, C_COLS 2 at 225 MHz, GT SerDes island; WNS +0.003 (true +0.093) [M F1299].
- First real models on the 8-card ring, bit-exact vs int sim: Llama-1B 398.83 [M F1341], 3B 169.1 with VPU4 [M F1325], 8B 97.44 tok/s [M F1343].
- Qwen3.6-27B real W8 TP8 9.85 tok/s, 36/36 greedy vs GPU golden [M F1318]. Open bug: ~1 AR in 300-1500 aborts, host-recovered [M F1310].
- PDI: a41a8e90 (talon p0 x8)
- silicon: ladder 52/52 x 8, BIST 16/16 dirs 0 errors [M F1308]
- Qwen3.8-27B: 19.37 tok/s, host levers [M F1395]
- DeepSeek: V4-Flash W8, 6 cards: 0.33 tok/s [M F1348]
- added / highlighted blocks: HBMS-AR <-> ring bridge; 8-card ring 17>3d>50>63>cf>e1>99>bd
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; QSFP28 Aurora ring core (cages 0 / 3) [4 x 25.78 Gb/s per direction, pace 8 + heal, DROP_RST; 99.84 Gb/s raw [M]]; HBMS-AR <-> ring bridge [async FIFO 225 -> 150 MHz, <=256-beat frames, ALLRED P=8]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; DeepSeek host-hybrid [6 cards, W8 experts from NVMe]

### DSpark v5 A6 (0x1267) (0xE4B71267) | 2026-10-21 | A-prime: 64 NMUs, mclk 300

- Target 2a line: 64 HBM NMUs + HBM-FOLD (2 per engine port), engine 225 / mclk 300; margin re-route passed QM-38: WNS +0.040 / WHS +0.013 [M F1383].
- SETBACK on talon: G0 8/8 PASS, but Llama-3B/8B and Qwen3.6 output wrong from token 0; fixed-lane data errors on engine ports 2/9/11 [M F1392].
- Same lane-fault class as 0x126D [M F1394]; the line became exact only at hbms_clk 150 / mclk 200 (0x1317) [M F1465].
- PDI: 252dd6b5 (mg_ds5r_V3p80)
- state: QM-38 PASS, silicon FAIL [M F1392]
- clocks: engine 225 / mclk 300 MHz
- added / highlighted blocks: HBM ports x32 + HBM-FOLD; Lane fault at mclk 300
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; HBM ports x32 + HBM-FOLD [64 NMUs, 2 per engine port]; QSFP28 Aurora ring core (cages 0 / 3) [4 x 25.78 Gb/s per direction, pace 8 + heal, DROP_RST; 99.84 Gb/s raw [M]]; HBMS-AR <-> ring bridge [async FIFO 225 -> 150 MHz, <=256-beat frames, ALLRED P=8]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; Lane fault at mclk 300 [output wrong from token 0]

### R2 done-right @225 (0x126D) (0xE4B7126D) | 2026-10-21 | done-right + on-card MX, mclk 300

- Done-right arm B (QWMX): adds the on-card Qwen MX engine (MXE0 / MXE1) to the 64-NMU HBM-FOLD target; QM-38 cleared with the reset-release MCP [M F1391].
- SETBACK: port/lane data errors on ports 7/12; corner-prohibit re-places cleared the lanes, but output stayed wrong from token 0 [M F1394, F1548].
- RCA: the exact build RTL == ISA emulator in xsim, so the fault is below the RTL [M xsim F1414]; fix = the 150 / 200 MHz derate (0x1317).
- PDI: 42226d20; re-places CPW 9a258951, CP2BS2 4e017b3f
- tok/s: 41-106 tok/s runs had wrong output: refuted [F1546-F1548]
- clocks: hbms_clk 225 / mclk 300 MHz
- added / highlighted blocks: MX engine: on-card Qwen mixer (MXE0 / MXE1); Lane fault at mclk 300
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; HBM ports x32 + HBM-FOLD [64 NMUs, 2 per engine port]; MX engine: on-card Qwen mixer (MXE0 / MXE1) [rtl_sd6: 288 DeltaNet slots (2 x 144), PM 524,288 words, ROM 512]; QSFP28 Aurora ring core (cages 0 / 3) [4 x 25.78 Gb/s per direction, pace 8 + heal, DROP_RST; 99.84 Gb/s raw [M]]; HBMS-AR <-> ring bridge [async FIFO 225 -> 150 MHz, <=256-beat frames, ALLRED P=8]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; Lane fault at mclk 300 [output wrong from token 0]

### v8 + ARQ ring fix (0x130E) (0xE4B7130E) | 2026-10-22 | B2r: ring link fix, DeepSeek host-hybrid | 169.19 tok/s [M] (Llama-3.2-3B TP8, 8 cards)

- v8 + ring link fix (FIX63 / ARQ) for the err 0x6 AR stalls: BIST 16/16 0 errors, ARQ counters 0, AR retries 0 [M F1407].
- Bit-exact batch-1: Llama-1B 413.66-422.48, 3B 169.19, 8B 101.23 tok/s [M F1407, F1408]; Qwen3.8-27B 16.62-19.08 tok/s, exact 4/4 [M F1409].
- DeepSeek-V4-Flash W8 host-hybrid on 6 cards, experts streamed from NVMe: 0.35 tok/s, tokens == the v8 run [M F1452].
- PDI: 433cf603 (talon p0 x8)
- timing: G-MARGIN PASS, true hbms_clk +0.090 [M F1402]
- demo: 1B 365 / 3B 166 / 8B 99.9 tok/s, live switching [M F1456]
- added / highlighted blocks: QSFP28 Aurora ring + ARQ link fix; DeepSeek host-hybrid
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; QSFP28 Aurora ring + ARQ link fix [FIX63 / ARQ; BIST 16/16 0 errors, 0 AR retries [M F1407]]; HBMS-AR <-> ring bridge [async FIFO 225 -> 150 MHz, <=256-beat frames, ALLRED P=8]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; DeepSeek host-hybrid [6 cards, W8 experts from NVMe]

### URE drafter pair (0x1306 / 0x1307) (0xE4B71306) | 2026-10-23 | 2-card URE drafter, Llama + Qwen

- Unified URE drafter on 2 V80s: HEAD 0x1306 HA_V2 on e1, TAIL 0x1307 T75_V3 on cf; the host relays drafts and verify results to the target cards.
- Silicon PASS on talon: HEAD 29 files byte-for-byte + NONZERO 32/32, TAIL 18/18 [M F1458].
- Draft step 3.89 ms overlapped / 6.27 ms serial vs GPU EAGLE3 3.555 ms [D F1473]; one URE session per boot [M F1473].
- PDIs: HEAD 2d85a008, TAIL fca9622b (talon e1 / cf p0)
- tok/s: drafter only, no end-to-end tok/s yet
- added / highlighted blocks: URE drafter HEAD (card e1); URE drafter TAIL (card cf)
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; URE drafter HEAD (card e1) [0x1306 (Llama eps) / 0x1319]; URE drafter TAIL (card cf) [0x1307; host relay link]

### CHAIN4 4+4 split (0x1314) (0xE4B71314) | 2026-10-23 | 4 target + 4 draft split, closed

- 0x130E + RING_CFG[6] open-chain all-reduce: two TP4 groups at once on one 8-ring; xsim 16/16 and non-BEH netlist gate PASS [M F1406, F1413].
- SETBACK: on talon every Llama 1B/3B/8B prompt is wrong from token 0 and Qwen exactness FAILs [M F1444]; the corner-DSP re-close 0x1314-CP is also not exact [M F1536].
- Speeds of wrong outputs are refuted [F1536]. Line closed; the chain moved to the exact done-right lineage as 0x1318 [M F1479].
- PDIs: b1 34c8ed05, CP 908ee653
- state: NOT exact on silicon, line closed [M F1479]
- RCA: 2026-10-22/SPLIT_4P4/RCA_1314_SILICON.md
- added / highlighted blocks: 4+4 split: two TP4 chains; Not bit-exact on silicon
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM port engines x32 [weight FIFO TH 32 beats; 225 MHz]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks [hbms_clk 225 MHz (0x1233: 200) · usr 50 MHz]; QSFP28 Aurora ring + ARQ link fix [FIX63 / ARQ; BIST 16/16 0 errors, 0 AR retries [M F1407]]; HBMS-AR <-> ring bridge [async FIFO 225 -> 150 MHz, <=256-beat frames, ALLRED P=8]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; 4+4 split: two TP4 chains [RING_CFG[6] on one 8-ring]; Not bit-exact on silicon [wrong from token 0; closed]

### DRM done-right (0x1317) (0xE4B71317) | 2026-10-23 | done-right at 150 / 200 MHz, talon demo image | 135 tok/s [M] (Llama-3.2-3B TP8, 8 cards)

- The 0x126D done-right RTL derated to hbms_clk 150 / mclk 200 (NoC read bw 6400): the first done-right image exact end to end on silicon [M F1465].
- Bit-exact batch-1 on 8 cards: Llama-1B 345, 3B 135, 8B 77.8 tok/s [M F1465-F1467]; Qwen3.8-27B via the on-card MX path 33.02 tok/s [M F1470].
- The talon demo image: PDI d76ab1b1 on talon p0 10-23 to 10-26; QM-38 passed as diagnostic-only, with a waiver on the async mclk recovery path (+0.006 ns) [M F1459].
- PDI: d76ab1b1 (buildz05 arm z05)
- timing: true hbms_clk +0.093, mclk +0.086; async mclk +0.006 waived [M F1459]
- demo v4: Qwen 32.48 / 33.81, Llama-8B 76.04 tok/s, 0 AR retries [M F1471]
- Llama runs: short 15-24 token selftests [M F1465-F1467]
- limit: Qwen spec decode (M>1) NOT valid: DSP58 lf_wa [M xsim F1532]
- added / highlighted blocks: Clocks (done-right); Ring all-reduce on card (done-right lineage); MX engine: on-card Qwen mixer (MXE0 / MXE1)
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM ports x32 + HBM-FOLD [64 NMUs, 2 per engine port]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks (done-right) [hbms_clk 150 / hbms_mclk 200 MHz]; MX engine: on-card Qwen mixer (MXE0 / MXE1) [rtl_sd6: 288 DeltaNet slots (2 x 144), PM 524,288 words, ROM 512]; QSFP28 Aurora ring + ARQ link fix [FIX63 / ARQ; BIST 16/16 0 errors, 0 AR retries [M F1407]]; Ring all-reduce on card (done-right lineage) [HBMS-AR <-> ring at 150 / 200 MHz; TP8 exact [M F1465]]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; Qwen spec decode invalid [M>1 verify: DSP58 lf_wa bug]

### DR_CHAIN6 6+2 (0x1318) (0xE4B71318) | 2026-10-24 | done-right + 6-card open chain

- 0x1317 + RING_CFG[6] open chain for 6 target + 2 URE: chain 63-50-3d-17-bd-99 (ranks 0-5), e1 / cf free for the drafter pair; 8-card mode == 0x1317.
- Gates: chain netlist 19/19 [M xsim F1480]; TP6 programs == 1-card int sim for Llama 1B/3B/8B [M F1482]; G-MARGIN PASS, async mclk +0.031, no waiver [M F1486].
- Not yet on silicon (talon down). Projected 6+2 Llama-3B as built: 106 tok/s, below TP8 target-only [P F1499].
- PDI: b1 4c5d32e3, QM PASS; b2 5cf01a17 spare [M F1514]
- Qwen TP6: does not fit: 288 of 384 DeltaNet slots [D F1492]
- limit: Qwen spec decode (M>1) NOT valid [M xsim F1532]
- added / highlighted blocks: Open chain: 6 target cards
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; GEMV engine [MD=4 activation columns; W8 / W3 / W2 per op]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM ports x32 + HBM-FOLD [64 NMUs, 2 per engine port]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks (done-right) [hbms_clk 150 / hbms_mclk 200 MHz]; MX engine: on-card Qwen mixer (MXE0 / MXE1) [rtl_sd6: 288 DeltaNet slots (2 x 144), PM 524,288 words, ROM 512]; QSFP28 Aurora ring + ARQ link fix [FIX63 / ARQ; BIST 16/16 0 errors, 0 AR retries [M F1407]]; Ring all-reduce on card (done-right lineage) [HBMS-AR <-> ring at 150 / 200 MHz; TP8 exact [M F1465]]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; Open chain: 6 target cards [63-50-3d-17-bd-99, RING_CFG[6]]; URE drafter HEAD (card e1) [0x1306 (Llama eps) / 0x1319]; URE drafter TAIL (card cf) [0x1307; host relay link]; Qwen spec decode invalid [M>1 verify: DSP58 lf_wa bug]

### URE HEAD eps fix (0x1319) (0xE4B71319) | 2026-10-24 | URE HEAD for Qwen

- URE HEAD with the Qwen eps 1e-6 + e_norm wrap / saturation fixes; pairs with TAIL 0x1307 on cf.
- HG_V2 PDI cc4d3c09: true WNS +0.045 / WHS +0.030 [M F1503]; non-BEH netlist gate 49/49 bit-exact to RTL [M xsim F1506].
- QM PASS, staged on talon; not yet on silicon (talon down).
- PDI: HG_V2 cc4d3c09 (preferred)
- backup: HH_V3 ab237555, WNS +0.041 [M F1503]
- added / highlighted blocks: URE drafter HEAD (card e1)
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; URE drafter HEAD (card e1) [0x1306 (Llama eps) / 0x1319]; URE drafter TAIL (card cf) [0x1307; host relay link]

### DRC6X + PFX + MR3 (0x131A) (0xE4B7131A) | 2026-10-25 | 6+2 with PFX all-reduce + MR3

- 0x1318 + PFX chain all-reduce (RING_CFG[7]) + MR3 multi-row GEMV (DSP58 INT8 3-lane, fold F3): the two RTL levers of the 6+2 projection.
- Gates: ring PFX on / off 57/57 [M xsim F1509], full core 19/19 [M xsim F1510], MR3 non-BEH netlist 4/4 [M F1513]; arm C G-MARGIN PASS, async mclk +0.044 [M F1526].
- Flashed on talon 10-26 (T2 relay), untested: talon down. Projected 6+2 Llama-3B at M=4: 165 tok/s = 1.27x TP8 0x1317 [P F1512].
- PDI: arm C accbfdd5, QM PASS (arm A 18a4a902 backup)
- MR3: K-padded M=4 1,841 vs M=1 1,777 cycles [M xsim F1498]
- limit: Qwen spec decode (M>1) NOT valid [M xsim F1532]
- added / highlighted blocks: PFX chain all-reduce (RING_CFG[7]); GEMV engine + MR3
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM ports x32 + HBM-FOLD [64 NMUs, 2 per engine port]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks (done-right) [hbms_clk 150 / hbms_mclk 200 MHz]; MX engine: on-card Qwen mixer (MXE0 / MXE1) [rtl_sd6: 288 DeltaNet slots (2 x 144), PM 524,288 words, ROM 512]; QSFP28 Aurora ring + ARQ link fix [FIX63 / ARQ; BIST 16/16 0 errors, 0 AR retries [M F1407]]; Ring all-reduce on card (done-right lineage) [HBMS-AR <-> ring at 150 / 200 MHz; TP8 exact [M F1465]]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; GEMV engine + MR3 [multi-row verify: DSP58 INT8 3-lane MAC, fold F3, M<=4]; PFX chain all-reduce (RING_CFG[7]) [prefix/suffix per link; LEN 6144 6,314 -> 2,682 cyc [M xsim F1505]]; Open chain: 6 target cards [63-50-3d-17-bd-99, RING_CFG[6]]; URE drafter HEAD (card e1) [0x1306 (Llama eps) / 0x1319]; URE drafter TAIL (card cf) [0x1307; host relay link]; Qwen spec decode invalid [M>1 verify: DSP58 lf_wa bug]

### UNIFIED6 (0x131B) (0xE4B7131B) | 2026-10-26 | Llama + Qwen TP6, DSP58 fix

- One 6-card target image for Llama and Qwen3.8-27B TP6: 0x131A + the TP6 MX engine, 2 x 192 = 384 DeltaNet slots per card (needed 384) [D F1517].
- DSP58 lf_wa fix (log address as shift-add): MXE netlist gate 16/16 + REDs 3/3 [M xsim F1565]; arm A2 PDI 152f73e0 G-MARGIN PASS after the reset-net re-route, async mclk +0.056 [M F1569].
- Not yet on silicon. Projected 6+2 Llama-3B at M=4: 165 tok/s [P F1512, proj_out.md row D]; DeepSeek Plan A on 6 cards 0.6 tok/s [P F1523].
- PDI: A2 152f73e0, QM PASS (no waiver)
- Qwen: TP6 only on this image; Qwen TP8 stays on 0x1317
- DeepSeek W8: needs 8.54 cards, does not fit 6 [D F1518]
- added / highlighted blocks: MX engine TP6 (MXE0 / MXE1, rtl_t6)
- blocks present: Host serve_card.py / serve_pp.py [token id in, argmax out]; QDMA H2C / C2H [weights load: 1.48 GiB in 1.8 s [M]]; MMIO BAR0: CSR, START, stats [1.37 us per read [M]]; PCIe Gen5 x8; Descriptor sequencer [256-op table, BR + LREL layer loop]; Spec-decode ISA (r7-r9) [VROWS ATTN.TREE TOPK TACCEPT GATHER PSTEP]; CAB activation + KV window [int8 activations; KV 27-32 pos (KV_MODE 0)]; Stat BRAM [per-op / per-port cycles (stall_tot 3.9% [M])]; VPU [RMSNORM ROPE SILUMUL RESADD ARGMAX; VPU4 (-3% [M])]; Attention [MQ, ATTN_MODE 0/1, KV_MODE 0/1]; HBM ports x32 + HBM-FOLD [64 NMUs, 2 per engine port]; HBM2e 32 GB [NP32: ~230 GB/s streamed [M]  ·  spec 819 GB/s]; Clocks (done-right) [hbms_clk 150 / hbms_mclk 200 MHz]; QSFP28 Aurora ring + ARQ link fix [FIX63 / ARQ; BIST 16/16 0 errors, 0 AR retries [M F1407]]; Ring all-reduce on card (done-right lineage) [HBMS-AR <-> ring at 150 / 200 MHz; TP8 exact [M F1465]]; 8-card ring 17>3d>50>63>cf>e1>99>bd [TP8 all-reduce on card]; GEMV engine + MR3 [multi-row verify: DSP58 INT8 3-lane MAC, fold F3, M<=4]; MX engine TP6 (MXE0 / MXE1, rtl_t6) [384 DeltaNet slots (2 x 192), PM 1,048,576 words, ROM 1,024]; PFX chain all-reduce (RING_CFG[7]) [prefix/suffix per link; LEN 6144 6,314 -> 2,682 cyc [M xsim F1505]]; Open chain: 6 target cards [63-50-3d-17-bd-99, RING_CFG[6]]; URE drafter HEAD (card e1) [0x1306 (Llama eps) / 0x1319]; URE drafter TAIL (card cf) [0x1307; host relay link]; DeepSeek host-hybrid [6 cards, W8 experts from NVMe]

## Track 3: Constant-folded compute planes

### TUNIT_0x1258 s2 (0xE4B71258) | 2026-10-14 | one-card correctness unit

- TriLM-99M layer 0: all 7 ternary projections folded into LUTRAM on one V80.
- Silicon: 0 mismatches / 10,240, 266 cycles per projection [M] (box1 card 42).
- Full model on one card reloads weights per layer over MMIO: ~0.47 tok/s [P]; the fold itself is ~1,677 tok/s [D].
- timing: WNS 0.587 / WHS 0.010 ns [M]
- silicon: 0 / 10,240 mismatches [M]
- added / highlighted blocks: Folded ternary planes (LUTRAM); Neuron adder trees
- blocks present: Host harness [embedding, norms, attention, LM head]; AXI-Lite MMIO [weight reload per layer (1.84 s/token) [P]]; PCIe; int8 activation bus [x in, 8-bit]; Folded ternary planes (LUTRAM) [7 projections (q k v o gate up down), 3.01 M weights {-1,0,+1} as LUT constants]; Neuron adder trees [266 cycles / projection @ 50 MHz [M]]; Neuron read-back [3.7 us each over MMIO]; Clocks [usr 50 MHz (TUNIT) · 150 MHz (FOLDRING8)]

### FOLDRING8 s3/s5 (0x1259 / 0x125A) (0xE4B71259) | 2026-10-15 | stopped: URAM infeasible | 4550 tok/s [D] (TriLM-99M, 8 cards, xsim)

- 8-card weight-resident TriLM ring; xsim G3-LINK: 6/6 tokens bit-exact, 32,985 cycles/token at 150 MHz [M xsim].
- SETBACK: synthesis found ram_style=ultra infeasible for K/V; ~50 Mbit fell back to LUTRAM. Stopped on the user's order.
- ~4,550 tok/s batch-1 [D from xsim cycles].
- state: STOPPED (invalid)
- xsim: 6/6 tokens bit-exact [M xsim]
- added / highlighted blocks: URAM K/V banks
- blocks present: Host harness [embedding, norms, attention, LM head]; PCIe; int8 activation bus [x in, 8-bit]; Folded ternary planes (LUTRAM) [7 projections (q k v o gate up down), 3.01 M weights {-1,0,+1} as LUT constants]; Neuron adder trees [266 cycles / projection @ 50 MHz [M]]; All layers weight-resident [no reload, no weight memory traffic]; URAM K/V banks [explicit URAM banks (s7); s3/s5 fell back to ~50 Mbit LUTRAM]; Clocks [usr 50 MHz (TUNIT) · 150 MHz (FOLDRING8)]; QSFP28 ring (talon b8 PHY) [8 cards, per-card clocks +-1000 ppm, 0 link errors in xsim [M xsim]]; 8-card layer pipeline [TriLM-99M across 8 V80s]

### FOLDRING8_fast s7 (0x125A) (0xE4B7125A) | 2026-10-15 | rebuild, explicit URAM | 4550 tok/s [D] (TriLM-99M, 8 cards, xsim)

- K/V rewritten as explicit URAM banks, with a synth-feasibility gate before the PDI.
- Building on olegvm3 (fr8_build_fast.sh s7fast, 150 MHz).
- First end-to-end constant-fold tok/s on silicon when it lands.
- state: BUILDING (olegvm3)
- added / highlighted blocks: URAM K/V banks; QSFP28 ring (talon b8 PHY)
- blocks present: Host harness [embedding, norms, attention, LM head]; PCIe; int8 activation bus [x in, 8-bit]; Folded ternary planes (LUTRAM) [7 projections (q k v o gate up down), 3.01 M weights {-1,0,+1} as LUT constants]; Neuron adder trees [266 cycles / projection @ 50 MHz [M]]; All layers weight-resident [no reload, no weight memory traffic]; URAM K/V banks [explicit URAM banks (s7); s3/s5 fell back to ~50 Mbit LUTRAM]; Clocks [usr 50 MHz (TUNIT) · 150 MHz (FOLDRING8)]; QSFP28 ring (talon b8 PHY) [8 cards, per-card clocks +-1000 ppm, 0 link errors in xsim [M xsim]]; 8-card layer pipeline [TriLM-99M across 8 V80s]
