Companion to the week-40 architecture report · 2026-10-05

EAGLE4 Roadmap Review Pack

What to read before and during the roadmap session, grouped by the decision each item feeds. Site pages are public links on this research site. Lab files are internal paths on the build server; we can publish any of them as a page on request. Order within each group is the suggested reading order.

0 · Start here

  1. EAGLE4 V80 Architecture (week 40): block diagram, 17 milestones, widths and memories per block, utilisation and timing. The baseline for every discussion below.
  2. Companion chip product page: the product framing (1.33× vanilla H200 on HumanEval, 331 tok/s MT-Bench, measured).
  3. RTL onboarding: repo layout, build and gate flow, for anyone who will read RTL with us.
  4. Lab file: full audit with evidence excerpts /shared/rnd_hq/2026-10-05/tracks/arch_audit/ARCH_AUDIT.md and evidence/.

1 · Where the time goes today (per-level budget)

  1. Lab file CEILING_AND_PIVOT.md: what limits batch-1 tok/s on a 2-card pair and where the ceiling sits. /shared/rnd_hq/2026-10-01/tracks/ceiling/
  2. Lab file HEADROOM_AUDIT.md: per-stage timing headroom (fmax, slack, failing-endpoint classes). /shared/rnd_hq/2026-10-01/tracks/headroom/
  3. Lab file VERIFY_COST.md: GPU target-verify cost per level, the floor no FPGA work can remove. /shared/rnd_hq/2026-10-01/tracks/verify_cost/
  4. Whole-drafter fmax census and dual-clock array: why the drafter runs at 66/75 MHz with a 2:1 array clock.
Discuss Is the remaining gap clock-bound (fmax of the drafter fabric) or latency-bound (host relay, verify)? Where would a hardened block or an ASIC change the answer?

2 · Compute array and memory

  1. Array architecture: weights resident per lane, cross-die activation broadcast (123 of 128 lanes fed from another SLR).
  2. Weight residency: 41.58 MiB on-chip per card, about 5× across 4 cards.
  3. Systolic / LUT-GEMM study: LUT reduction vs clock.
  4. Lab file BIG_ARRAY.md: wider arrays (K_PAR32, HX2/HX3), and why they failed timing or exactness. /shared/rnd_hq/2026-10-02/tracks/big_array/
  5. Lab file hx2q_timing/REPORT.md: cross-SLR broadcast as the binding constraint for a 32-wide array. /shared/rnd_hq/2026-10-03/tracks/hx2q_timing/
Discuss Activation broadcast across SLRs limits array width. How would your fabric or NoC (or a 2.5D/3D memory stack) change the weight-resident vs streamed trade-off?

3 · Scaling out: 4 cards and card-to-card links

  1. Lab file ARCH_4CARD.md: the 4-card single-stream options compared. /shared/rnd_hq/2026-10-03/tracks/four_card_pipeline/
  2. Lab files LINK_TOPOLOGIES.md, LINK_DMA_V2.md, CAPACITY_ACCEPT.md: link options and capacity vs accept. /shared/rnd_hq/2026-10-01/tracks/four_fpga_research/
  3. Lab file QSFP_MULTIPORT.md: QSFP28 Aurora 4×25.78 Gb/s, multi-port topologies. /shared/rnd_hq/2026-10-02/tracks/qsfp_multiport/
  4. Lab files CL4_SPEC.md, CL4_INTEG_REPORT.md, CABLING_CL4.md: the CL4 ring (HEAD↔HEAD merge), projected 366–368 tok/s. /shared/rnd_hq/2026-10-03/tracks/four_card_pipeline/cl4/
  5. Lab file D2_CARD_LOOP.md: moving the level loop onto the cards. /shared/rnd_hq/2026-10-01/tracks/card_loop/
  6. Lab file HOSTLESS_RING.md: removing the host from the data path. /shared/rnd_hq/2026-10-02/tracks/hostless_ring/
  7. Lab file sibxe_silicon/RCA.md: first 4-card silicon run (50.00 vs 96.95 tok/s, KV-bound) and what it taught. /shared/rnd_hq/2026-10-03/tracks/sibxe_silicon/
Discuss Every 4-card gain so far depends on keeping KV and control off the host. What is the right interconnect (QSFP, PCIe P2P, a custom die-to-die link) for a product version?

4 · Platform and host path

  1. Lab file slowdown_4card/RCA.md: SMI storm on the lab chassis, 60–78% of CPU samples in SMM. Explains why host latency dominates. /shared/rnd_hq/2026-10-03/tracks/slowdown_4card/
  2. Lab file BIOS_CHECKLIST_BOX1.md: the platform settings that matter. /shared/rnd_hq/2026-10-03/
  3. Lab file NATIVE_RELAY.md: compiled host relay replacing Python glue. /shared/rnd_hq/2026-10-02/tracks/native_relay/
  4. Host overhead and host glue structure: CUDA-event profiling of the host path.
  5. Lab file DRAFT_DURING_VERIFY.md: overlapping drafting with GPU verify. /shared/rnd_hq/2026-10-01/tracks/draft_overlap/
Discuss A host CPU in the loop costs milliseconds per level. Which pieces belong in a companion chip's own control processor?

5 · Bigger models (Llama-3.1-8B, Qwen 27B)

  1. Lab file shape_generic/REPORT.md: runtime shape table, URAM budget per model, 4-card tensor-parallel split, projections (8B ~303, Qwen-27B ~172 tok/s, projected). /shared/rnd_hq/2026-10-05/tracks/shape_generic/
  2. Lab file T1_KIT.md: drafter retraining recipe (reduced draft vocab). /shared/rnd_hq/2026-10-01/tracks/drafter_retrain/
  3. Accept vs tree width/depth on GPU and draft KV lifecycle: the algorithm-side levers.
Discuss A 27B target with hybrid DeltaNet layers moves the bottleneck to GPU verify. Should the companion chip target the drafter only, or part of verify too?

6 · Ideas backlog and history

  1. Lab file INNOVATION_BOARD.md: every idea tried or queued, with verdicts. /shared/rnd_hq/2026-10-02/
  2. Lab file FOUR_CARD_INNOVATIONS.md. /shared/rnd_hq/2026-10-02/tracks/four_card_innovations/
  3. Drafter journey, week review 09-24..27, progress 09-27..29: how we got from 153 to 339 tok/s.
  4. Ledgers: findings.csv (every measured/derived/projected number with evidence), branches.csv, images.csv. /shared/rnd_hq/hq_daily/