ACU — Attention Compute Unit

The ACU is the decode-attention datapath of the LonghornSilicon (Lambda) LLM inference accelerator: Q·Kᵀ → softmax → P·V, with a per-tile INT8/FP16 precision gate. It lives at src/blocks/acu/ (reorg 2026-07; see docs/REORG_NOTES.md) and is a multi-block unit, so it splits into three self-contained sub-blocks — each carrying the canonical block layout (sw/ rtl/ pdk/ docs/ research/, per SOP §5.1) and mirroring on its own:

Sub-block	What it is	Mirror
mate/	MatE — the matmul engine: mate_qkt (Q·Kᵀ scoring), mate_pv (INT8 P·V), mate_pv_fp16 (FP16 P·V) + the 8×8 MAC-array reference model	lambda-mate
vecu/	VecU — the decode online-softmax slice (vecu_softmax, exp-LUT + fp32 accum) plus RoPE / RMSNorm tiles	lambda-vecu
precision_controller/	Per-tile INT8-vs-FP16 gate: the pre-softmax ratio test `max(	S

The assembled unit mirrors to lambda-acu (umbrella); each sub-block also mirrors on its own.

Where things are

• RTL + tb → src/blocks/acu/<sub>/rtl/ (SystemVerilog; the shared multi-tile synth/EDA harness lives in src/blocks/acu/mate/rtl/).
• Reference models → src/blocks/acu/<sub>/sw/reference_model/.
• Hardening → src/blocks/acu/<sub>/pdk/<target>/: sky130/openlane/ (flagship dev/proof), asap7/orfs/ (predictive 7nm bracket), gf180/librelane/ (SSCS Chipathon 2026 shuttle configs).
• Design notes / ISA → src/blocks/acu/<sub>/docs/; ACU-wide docs → src/blocks/acu/docs/ (incl. the legacy full overview docs/acu_overview.md — its paths are pre-reorg).
• The RL research that birthed the ACU (Adaptive Precision Attention policy evolution) → research/apa-precision-policy/ at the repo root (chip-wide research).

Status

Honest per-macro, per-PDK sign-off from docs/PROGRESS.md (generated from committed *metrics.json; sign-off definitions in docs/REVISION_SYNC_SOP.md §5.2). Sign-off is per flow, not per block.

• sky130 (OpenLane) — signed-off: mate_pv (75,660 µm² / 71 MHz), mate_pv_fp16 (221,987 µm² / 12 MHz), mate_qkt (198,901 µm² / 12 MHz), vecu rope (277,632 µm² / 10 MHz) and vecu_softmax (726,020 µm² / 10 MHz), precision_controller (8,642 µm² / 80 MHz). rmsnorm on sky130 is config-only.
• gf180 (LibreLane) — signed-off: vecu rope (432,140 µm² / 4 MHz) and rmsnorm (1,095,180 µm² / 4 MHz). Everything else on gf180 (mate_*, vecu_softmax, precision_controller) is config-only (declared, not run) — GF180 is the tape-out PDK, assembled via the chip-level padring in chip/pdk/gf180/.
• asap7 (ORFS) — route-clean (NOT full sign-off): mate_pv (2.0 GHz), mate_pv_fp16 (286 MHz), precision_controller (1.18 GHz). ASAP7 runs no Magic-DRC and no LVS — a predictive 7nm bracket only.

Known gotchas

• "Signed off" is per flow. Do not read "signed off on sky130" as "signed off everywhere" — most gf180 macros are config-only and every asap7 run is route-clean, not signed off.
• ASAP7 ORFS is 4×-drawn — areas read 16× too large unless de-scaled (confirm SITE 0.054×0.270).
• FP16 can't be bit-exact to numpy @ (BLAS pairwise sum ≠ sequential MAC). Verify FP16 RTL vs a sequential-fp32 golden; tolerance vs numpy rel_err < 5e-3.
• The vecu_softmax exp-LUT carries ~2% error vs exact softmax (64-entry linear interp over [-16,0]); cosim tolerances are set FROM it, not tighter.
• Long combinational fp paths won't close at the slow corner — pipeline them (decode is latency-tolerant).
• precision_controller.sv is byte-identical to the out-of-band attention-compute-unit repo — a one-time manual import with no auto-sync, so either side can drift silently. Future changes originate here (SOP §7).
• The 16nm area/speed projections do not reconcile between Sky130-scaled and ASAP7-derived (~5× apart); treat 16nm as a range pending a real Cadence 16FFC run. See docs/acu_overview.md.

Branch model

main is a clean scaffold — structure, docs, pdk/ configs, Python golden models, results/ — but no .sv/.v RTL. The RTL lives on the rev0 revision branch: contributors PR into rev0, a lead blesses (RTL reviewed, sign-off reproduced) and merges upstream to main. To view/work on RTL: git checkout rev0. Full model: docs/REVISION_SYNC_SOP.md §6a.

Before you touch this block

Read AGENTS.md (front door), the sub-block DECISIONS.md, and the sub-block README's ## Known gotchas. Follow the lab-notebook rule: docs travel with code in the same commit.