Weekly report · week 40 · 2026-10-05 · speculative-decoding drafter on AMD Alveo V80

EAGLE4 V80 Architecture

The EAGLE drafter for Llama-3.2-3B, split across V80 cards (HEAD and TAIL) with an H200 doing target verification. Pick a milestone to see which hardware it added. Every number is labelled M measured from RTL, a built wrapper or a routed report, D derived, or P projected. Items not checked are marked unverified.

Architecture timeline

Added or changed at this milestone HEAD card TAIL card GPU / host / PCIe Card-to-card link / 4-card Not yet present

Block details

Select a block in the diagram.

Batch-1 throughput by milestone

Filled dots are measured on silicon, hollow dots are projections. The 4-card SIBX-E point (50.0) ran on a chassis slowed 2–3.5× by an SMI storm, so it is not comparable with the record pair.

Full architecture audit (current state)