PR-G: Formal Industrial Benchmark v2 measurement on the qualified CUDA host

[telleroutlook]

Purpose

Execute the formal Industrial Benchmark v2 measurement on the already-qualified controlled NVIDIA/CUDA host.

This issue creates measurement evidence only.

Publication, README editing, and final repository integration remain maintainer-owned.

---

Hard prerequisites

Maintainer decision (2026-09-25): the formal commit is PRE-FROZEN above (a70d4ac76fc73e1f1e1099c7884db9bfdd4802d3) so this handoff needs no further maintainer input on the normal path. The DO NOT START UNTIL #56 PASS gate below still applies unchanged — if #56 exposes a source problem, the standing firewall invalidates this freeze and the SHA will be re-issued.

BLOCKED BY:
- #58
- #56

Execution order is strict:

#58 CLOSED
→ new frozen measurement commit
→ #56 PASS
→ #57 may begin

The commit used by #57 MUST be the same source authority qualified by #56,
unless a new source change occurs.

If any source change occurs after #56 PASS:
#56 qualification is invalidated
→ a new commit must be frozen
→ #56 must be rerun
→ only then may #57 begin.

Do not begin until all are true:

• Phase 2B.2 authority-repair issue is CLOSED.
• CUDA host qualification issue is PASS.
• The maintainer has frozen the exact formal measurement commit below.
• The exact Ibex fixture is available.
• No benchmark/protocol source changes are pending.

Formal commit:

FROZEN_FORMAL_MEASUREMENT_COMMIT:
a70d4ac76fc73e1f1e1099c7884db9bfdd4802d3

Host qualification issue:

#56

---

Authority firewall

During this formal run:

NO source edits
NO benchmark edits
NO dependency edits mid-run
NO parameter tuning after observing results
NO manual JSON editing
NO manual canonical-file copying
NO reuse of provisional numbers
NO best-of-N reporting

Any required source fix means:

STOP
→ new commit
→ new run identity
→ rerun host qualification
→ restart formal measurement from zero

---

Fixture authority

Use exactly:

Source:
SJTU-YONGFU-RESEARCH-GRP/PDB-Physical-Design-Database

Source commit:
9e1e3399b1b707f26fee853bce1ff91ab466ce24

GDS:
layout/sky130hd/ibex/ibex.gds

Layer:
66:44

Pixel:
1.0 nm/px

Record:

sha256sum "$GDS"

---

1. Verify exact checkout

cd OpenLithoHub

git checkout a70d4ac76fc73e1f1e1099c7884db9bfdd4802d3
git rev-parse HEAD
git status --porcelain

Required:

HEAD == frozen formal SHA
working tree == clean

---

2. Capture host/environment evidence

Before measurement:

mkdir -p measurement-logs

git rev-parse HEAD \
  | tee measurement-logs/git-head.txt

git status --porcelain \
  | tee measurement-logs/git-status-before.txt

sha256sum "$GDS" \
  | tee measurement-logs/fixture-sha256.txt

nvidia-smi -q \
  | tee measurement-logs/nvidia-smi-q.txt

pip freeze \
  | tee measurement-logs/pip-freeze.txt

Also record:

python - <<'PY' | tee measurement-logs/torch-environment.txt
import torch
print("torch:", torch.__version__)
print("torch CUDA:", torch.version.cuda)
print("CUDA available:", torch.cuda.is_available())
print("device count:", torch.cuda.device_count())
print("GPU:", torch.cuda.get_device_name(0))
print("cuDNN:", torch.backends.cudnn.version())
print("TF32 matmul:", torch.backends.cuda.matmul.allow_tf32)
print("TF32 cuDNN:", torch.backends.cudnn.allow_tf32)
PY

Operator logs are supporting context only. Harness-owned environment lock remains canonical.

---

3. Formal preflight

python scripts/preflight_industrial_v2.py \
  --gds "$GDS" \
  --device cuda:0 \
  | tee measurement-logs/preflight.txt

Required:

PREFLIGHT: PASS

Otherwise STOP.

---

4. Run the exact formal protocol

Run:

python benchmarks/industrial-v2/run_v2_benchmark.py \
  --tiers a,b,c \
  --gds "$GDS" \
  --device cuda:0 \
  --dtype fp32 \
  --repeats 5 \
  --warmup 2 \
  --batch 8 \
  --tile 1024 \
  --halo 64 \
  --layer 66:44 \
  --windows 4096,8192,16384,32768 \
  --hopkins-grid 1024 \
  --formal \
  2>&1 | tee measurement-logs/formal-run.txt

Do not alter these parameters after seeing results.

---

5. Required Tier A result

Every declared Tier A window must be present:

4096
8192
16384
32768

Each must satisfy exact indexed/reference semantic parity for:

• candidate rows;
• owned runs;
• run IDs;
• contributor object IDs;
• window raster;
• verification ownership metadata.

Required:

status = SUCCESS
correctness_witness_pass = true

Tier A performance may be weak. Weak speedup is not failure.

---

6. Required Tier B result

Every declared window must complete.

Required:

status = SUCCESS
correctness_witness_pass = true
timing_method = cuda_synchronized
repeat_count >= 5

Artifacts must contain:

CPU batch=1 witness
GPU batch=1
GPU batch=N
host peak RSS
GPU allocated peak
GPU reserved peak
fresh-process repeat statistics
median / p10 / p90 / n

A slow GPU result is valid evidence.

Do not rerun with different parameters to obtain a better number.

---

7. Required Tier C result

Required:

status = SUCCESS
correctness_witness_pass = true
finite_witness = true
cpu_deterministic_witness = true
timing_method = cuda_synchronized
repeat_count >= 5

Cold and warm timing must remain distinct.

The claim-bearing warm statistic must follow the protocol's repeat aggregation and must not use best-of-N selection.

---

8. Canonical family build

A successful formal run must report:

canonical_build_blockers = []

and the run workspace must contain the complete canonical family:

industrial-v2-index.json
industrial-v2-gpu-runtime.json
industrial-v2-hopkins.json
industrial-v2-run-config.json
industrial-v2-distribution-freeze.txt
manifest.json
SHA256SUMS.txt

Do not manually construct or edit any of these files.

---

9. Promote to a clean staging root

Use only the supported promotion command introduced by the pre-measurement repair issue.

Example:

python scripts/promote_industrial_v2_artifacts.py \
  --workspace "<RUN_WORKSPACE>" \
  --canonical-root "./industrial-v2-canonical-staging"

Required result:

exactly seven canonical members
no blockers

Never use manual cp as canonical promotion.

---

10. Verify the staged family

python scripts/verify_industrial_v2_artifacts.py \
  --canonical-root ./industrial-v2-canonical-staging

Required:

V2 VERIFIER: PASS

If verifier fails, STOP.

Do not edit artifacts to make verification pass.

---

11. Final integrity check

git rev-parse HEAD \
  | tee measurement-logs/git-head-after.txt

git status --porcelain \
  | tee measurement-logs/git-status-after.txt

Required:

same frozen commit
clean tracked tree

---

12. Prepare handoff bundle

Create two archives.

Canonical candidate family

tar -czf industrial-v2-canonical-candidate.tar.gz \
  industrial-v2-canonical-staging

Operator/context evidence

tar -czf industrial-v2-measurement-evidence.tar.gz \
  measurement-logs \
  <RUN_WORKSPACE>

Record:

sha256sum \
  industrial-v2-canonical-candidate.tar.gz \
  industrial-v2-measurement-evidence.tar.gz

Attach both archives or provide them to the maintainer through the agreed project channel.

---

Required completion report

Post a final comment containing:

Formal measurement commit:
Run identity:
Fixture SHA-256:
GPU:
Driver:
CUDA:
PyTorch:
cuDNN:

Tier A: PASS / FAIL
Tier B: PASS / FAIL
Tier C: PASS / FAIL
Canonical build blockers:
Verifier: PASS / FAIL

Canonical archive SHA-256:
Evidence archive SHA-256:

Do not summarize performance selectively.

---

Explicit non-goals

The GPU operator must NOT:

• edit README;
• edit README_zh;
• create marketing claims;
• decide which measurements become headlines;
• compare with commercial tools;
• weaken thresholds;
• change mandatory tiers;
• change repeat counts;
• create a new benchmark protocol;
• mutate Industrial Benchmark v1.1;
• mutate P-054/B04 authority.

Those tasks remain maintainer-owned.

---

Success state

Only the following closes this issue:

FORMAL CUDA MEASUREMENT COMPLETE
TIER A/B/C SUCCESS
CANONICAL FAMILY BUILT
V2 VERIFIER PASS
EVIDENCE HANDED OFF

After that, maintainers may open/execute the separate Phase 2C publication closure.