EAGLE4 2-bit drafter architectureone forward pass, drawn as a processor pipeline — left→right

⚠ RETRACTED FIGURE ON THIS PAGE — “G8 = 93% (93.4%) of the forward”. (banner added 2026-09-02; page text below is left standing, struck, not deleted)
The 93.4% figure was retracted 2026-07-31 — it describes a K_C=512, pre-WIDE xsim frame, not the shipped operating point. Its replacement, “G8 = 19.0%”, was itself refuted 2026-08-16.
Correct values, with their regimes: at the shipped operating point (md=2, K_C=64, FR-Spec 32K) G8 = 10.93% of AP-BUSY. At K_C=512 it is 49.4% — and K_C=512 is not the shipped point.
Mechanism. In dbg_observe.sv:306-313 phases [0..11] are each a disjoint top_state == N decode, but bit [12] G8 is g8_in_gather — a datapath liveness signal, not a state match — so it spans S_TREE ∪ S_G8 and double-counts the tree walk. Corrected: 108,873 − 46,260 (TREE) = 62,613 cyc. The tell was arithmetic: the phase shares summed to 107.15% of AP-BUSY (614,014 vs 573,054 cyc), and a disjoint decomposition cannot exceed 100%.
⚠ The subtrahend 46,260 is 0x1143-at-nt=256 ONLY. 0x115B reads 4,118 (nt=256) / 16,054 (nt=1002). Do not apply one image’s TREE to another image’s G8.
🔸 And the doctrine has moved again since. G5_81PCT_AUDIT_20260902.md establishes that at real serving context it is G5 attention — not G4/G9 — that is the top phase: G5 passes G4 at ctx≥19 and all five body GEMVs at ctx≥50, on the measured law G5(ctx) = 672·ctx + 1442. The G5 share itself is bounded 60–81%, and 76.4% after the nt=256/1002 splice correction — do not quote 81%.
Sources: reports/results/G8_COUNTER_DISJOINT_SPEC_20260902.md · G8_SHARE_RECORD_SWEEP_20260902.md · SILICON_0x1143_FULL_CHARACTERISATION_20260814.md · CLAUDE.md:200. This surface sweep: reports/results/G8_93PCT_SURFACE_SWEEP_20260902.md.
What this costs this page’s argument: this page ranked G8’s HBM gather as “the wall-clock bottleneck” of the forward. That ranking rested on the retracted share and no longer follows. At the shipped operating point the top phases were G4 QKV (23.2%) and G9 down (20.3%), and at real serving context it is G5 attention — G8 was never the wall.

Every stage of the drafter, in the order the hardware runs it. Each block's shape encodes its operation class; the label under the name is the matrix it computes [out×in]; the line between blocks is annotated with the vector width flowing; a small colored tag marks where the weights/data live (URAM / BRAM / HBM / host). Two distinct silicon engines do the heavy lifting — see the color of each block and the note below the diagram. Numbers are verified against reference/include/eagle4_gemm_abstract.h.

Operation shapes
GEMV (2-bit × INT16) RMSNorm attention / softmax SiLU (compute-special) top-K funnel gather (skew = data move)
Two silicon engines
2-bit body GEMV (Wallace CSA LUT array + 1 DSP58/lane) INT16×INT16 DSP58 attention (HEAD_PAR=6)
Where data lives
URAM BRAM HBM (off-die) host / GPU
AMD V80 (XCV80) DIE — on-chip, zero HBM traffic except G8 2-BIT GEMV ENGINE DSP ATTN BACKEND · TREE + G8 fc root fuse [3072×9216] fp16 host / GPU depth-0 root only 3072 G1 RMSNorm BRAM γ 3072 G4 QKV proj [9216×6144] URAM 2-bit 13.5 MiB BRAM scales 9216 Q|K|V G5 attn 24-head QKᵀ·softmax·A·V K-cache URAM V-cache BRAM 3072 G6 O-proj [3072×3072] URAM 2-bit 2.25 MiB R3 ⊕ 3072 G7g gate G7u up [8192×3072] ×2 URAM 2-bit 6+6 MiB 8192 ×2 G7p SiLU·up gate⊙σ(gate)·up 8192 G9 down proj [3072×8192] URAM 2-bit 6 MiB → for_reasoning (depth≥1) 3072 G11 RMSNorm BRAM γ 3072 GLM lmdown fp16 [256×3072] URAM/BRAM 1.5 MiB 256 G12 cheap lm-head [128256×256] URAM 2-bit 7.8 MiB 128,256 logits TREE top-K_C select on-chip · depth-N 512 cand G8 exact rescore gather + dot HBM fp16 lm-head 751 MiB (788 MB) gather 512×3072 512 scores → top-K_F=8 OUT 8 frontier

Block width is schematic (not to time-scale) — for the time-proportional silicon datapath see the silicon datapath page. G8's HBM gather (751 MiB) is the wall-clock bottleneck (56–93% of the forward) → NOT the bottleneck: G8 = 10.93% of AP-BUSY at the shipped operating point (md=2, K_C=64); 49.4% only at K_C=512, which is not shipped. The MAC count was never the issue either — but neither is the gather, at the point the image actually runs.

§ enginesTwo engines, two precisions, two silicon fabrics

2-bit body GEMV engine URAM-resident weights

Every body projection — G4, G6, G7g, G7u, G9, G12 — is a matrix–vector product of 2-bit GPTQ weights (group-size 128, fp16 per-group scales) against INT16 activations. Runs on a Wallace 3:2 CSA-tree LUT MAC array (u_array); DSP58 declines at 2-bit granularity, so the reduction is pure LUT. A single DSP58 per lane applies the Q1.15 dequant scale. LANES=128 output columns × K_PAR=16 reduction. All weights live in URAM — zero HBM traffic at compute time.

INT16×INT16 DSP58 attention engine separate fabric

Attention (G5) is a different precision — INT16×INT16 dot products (attn_score_dot) mapped onto DSP58, time-multiplexed across the 24 heads. HEAD_PAR=6 replicates the engine 6× for a 6.00× head-parallel speedup. K-cache lives in URAM, V-cache in BRAM; there is no weight matrix — the operands are the per-branch KV context. Different precision ⇒ different silicon, kept off the GEMV array.

Norm (G1/G11), residual adds, SiLU (G7p) and the TREE funnel are lightweight scalar/elementwise logic on neither heavy engine. fc (fp16 root fuse) runs off-FPGA on the host/GPU and only at depth-0; depth≥1 reuses the parent's for_reasoning and skips it.

§ weightsPer-operation compute & memory requirements

Sizes re-derived as params × bits / 8. 2-bit = params/4 bytes; fp16 = params×2 bytes. MiB = 1,048,576 bytes.

StageOperationMatrix [out×in]Dtype ParamsSizeResidentMAC / fwd
fcroot fuse (3×h → 1)[3072×9216] fp1628.31 M54.0 MiB host/GPU28.3 M*
G1RMSNorm (input emb)γ[3072] fp163,0726 KiB BRAM—
G4QKV projection GEMV[9216×6144] 2-bit56.62 M13.50 MiB URAM56.6 M
G5attention, 24-head (MHA)KV [ctx×24×128] INT16—∝ ctx K URAM V BRAM∝ ctx
G6O-projection GEMV[3072×3072] 2-bit9.44 M2.25 MiB URAM9.4 M
G7ggate GEMV[8192×3072] 2-bit25.17 M6.00 MiB URAM25.2 M
G7uup GEMV[8192×3072] 2-bit25.17 M6.00 MiB URAM25.2 M
G7pSiLU(gate)·upelementwise [8192] INT16—— FF/regs—
G9down GEMV[3072×8192] 2-bit25.17 M6.00 MiB URAM25.2 M
G11final RMSNormγ[3072] fp163,0726 KiB BRAM—
GLMlmdown low-rank proj[256×3072] fp160.79 M1.50 MiB URAM/BRAM0.8 M
G12cheap lm-head logits[128256×256] 2-bit32.83 M7.83 MiB URAM32.8 M
TREEtop-K_C survivor selectK_C=512 ← 128256 ——— on-chip—
G8exact rescore = gather + dot[128256×3072] fp16394.0 M751 MiB
(788 MB)
HBM1.57 M**
Total 2-bit body weight (G4+G6+G7g+G7u+G9+G12) 41.58 MiBURAM175.2 M‡

* fc is fp16 and runs off-FPGA (host/GPU), only at the depth-0 root — not a per-forward FPGA cost.
** G8 multiplies only the K_C=512 gathered rows (512×3072 = 1.57 M MAC), but must fetch 751 MiB (788 MB) of fp16 lm-head from HBM — it is memory-bound, and this gather (not the MAC) dominates wall-time at 56–93% of the forward → costs 10.93% of AP-BUSY at the shipped operating point (md=2, K_C=64), 49.4% at the unshipped K_C=512.
‡ Total body MAC/forward ≈ 175.2 M (G4+G6+G7g+G7u+G9+GLM+G12, attention excluded — it scales with context). The 2-bit body weights (≈ 41.6 MiB) fit entirely in URAM, so the only HBM traffic in the whole forward is G8's exact-rescore gather.