Summary

Wires GPU acceleration for the compression proof (data-availability-v2)
into the production prover and plumbs in — but does not enable — the GPU
path for the aggregation proof.

• Compression: auto-enabled whenever a CUDA device is reachable. Wall-clock on
  the reference host drops from ~4:40 (CPU) to ~2:10 per proof (3-proof
  batch average 2:10.19 on AWS g7e.8xlarge with one RTX PRO 6000 Blackwell).
• Aggregation: GPU code is wired (gpu/plonk2 PI/BW6/BN254, gpu/vortex PI
  MiMC + ring-SIS, gpu/quotient) but off by default. Operators must set
  LINEA_PROVER_GPU_AGGREGATION=1 to opt in. Production should leave it off
  for now.
• Controller: when launched on a GPU host, only compression jobs are
  accepted. Execution / aggregation / invalidity files are ignored even if
  the corresponding Enable* toggles are on, so a GPU host never falls back
  to a slow CPU path for non-compression work.

⚠️ DevOps must read this

1. Build flags

The CPU binary is unchanged. The GPU binary requires the cuda build tag and
links against the static libgnark_gpu.a produced from prover/gpu/cuda.

# CPU only
make bin/prover                    # GO_BUILD_TAGS=debug, no CUDA dependency

# GPU
make bin/prover-cuda               # GO_BUILD_TAGS=debug,cuda; needs libgnark_gpu.a
# equivalently:
make GO_BUILD_TAGS=debug,cuda bin/prover

bin/prover-cuda is a new make target. The static library
prover/gpu/cuda/build/libgnark_gpu.a must already exist before linking; the
CMake build is unchanged from the existing gpu/cuda source tree.

2. Host class per job type

3. Controller behavior on GPU hosts

cmd/controller checks gpu.HasDevice() at start. When true:

• It accepts only *-getZkBlobCompressionProof.json jobs.
• EnableExecution, EnableAggregation, EnableInvalidity from the config
  are ignored even if they are true.

This is intentional — running the CPU-paths on a GPU host with 32 vCPU would
be much slower than dispatching them to the existing CPU pool.

CPU controllers are unchanged: same Enable* semantics as today.

4. Required runtime env vars

Compression (GPU host):

GOMEMLIMIT=180GiB
GOGC=75
# nothing else; GPU is auto-detected

These two values are baked into the reference run and keep peak Go heap
usage at ~200 GiB on a 249 GiB host without thrashing the GC.

Aggregation (CPU host) — unchanged from origin/main today.

5. Required runtime resources

• GPU host (compression): 1× RTX PRO 6000 Blackwell, 96 GiB VRAM. Peak
  VRAM usage observed: ~80 GiB. Do not schedule another GPU process on
  the same card while a compression proof is in flight.
• Disk: the prover-assets 7.1.0/data-availability-v2/ directory must be
  present on the host. The canonical SRS is read once per process and
  benefits substantially from being in OS page cache; a freshly-booted host
  pays ~2 min of cold-cache cost on the first proof. Subsequent runs
  hit the table below.
• Memory: peak host RSS ~210 GiB (large because GPU pinned-memory
  staging buffers are reused across rounds and the gnark Go heap is
  intentionally large under GOMEMLIMIT=180GiB).

6. Compression reference numbers (3 sorted requests)

Average wall time: 2:10.19

Per-phase decomposition (from prover logs): solve 33 s → init GPU
instance ~19 s → MSM commit L,R,O ~4 s → build/iFFT/commit Z ~8 s →
quotient GPU ~25 s → MSM h₁,h₂,h₃ ~4 s → eval+linearize+open Z ~7 s →
batch opening ~4 s.

Raw artifacts under prover/reference-benchmarks/results/2026-05-08-g7e-8xlarge-gpu-compression-final/.

7. Proof-flow summary

Compression (GPU host)
  controller picks up *getZkBlobCompressionProof.json*
  └─ bin/prover prove ...
     ├─ LoadSetup (canonical SRS only — GPU path)         ~17s
     ├─ Solver (gnark constraint system)                  ~33s
     └─ gpu/plonk2/bls12377.GPUProve (BLS12-377 PlonK)   ~1:43

Aggregation (CPU host, unchanged)
  controller picks up *getZkAggregatedProof.json*
  └─ bin/prover prove ...
     ├─ makePiProof  → PI wizard + BLS12-377 PlonK   (CPU)
     ├─ makeBw6Proof → BW6-761 PlonK                 (CPU)
     └─ makeBn254Proof → BN254 emulation PlonK       (CPU)

Aggregation (GPU host, opt-in only — DO NOT ENABLE TODAY)
  Same flow + LINEA_PROVER_GPU_AGGREGATION=1 → all three Plonk phases on
  gpu/plonk2; PI Vortex MiMC and ring-SIS on gpu/vortex; quotient
  evaluation on gpu/quotient.

8. Rollback

The compression GPU path can be disabled at runtime by deploying the
non-cuda bin/prover (or by hiding the GPU device from the prover process,
e.g. CUDA_VISIBLE_DEVICES=""). No code change required — the prover falls
back to gnark's CPU PlonK prover.

The aggregation GPU path is off by default; nothing to roll back unless an
operator explicitly set LINEA_PROVER_GPU_AGGREGATION=1 (just unset it).

Test plan

• go build ./... (CPU)
• go build -tags cuda,debug ./... (GPU)
• go test ./gpu/plonk2/... -tags cuda,debug (per-curve correctness vs gnark CPU reference)
• go test ./gpu -tags cuda,debug (device singleton)
• go test ./circuits/... ./backend/aggregation/... ./cmd/controller/... (touched packages)
• End-to-end compression on provertestdata2 × 3 sorted requests, all valid
• Smoke run by devops on a staging GPU host (g7e.8xlarge or equivalent)
• Confirm the controller on a CPU host still accepts execution/aggregation jobs
• Confirm the controller on a GPU host rejects execution/aggregation jobs

🤖 Generated with Claude Code