Hey Logos builders,
I run a dedicated GPU proving backend (RTX 4090, CUDA 12.8, RISC Zero 3.0.6 / SP1 v6.x) and I’m offering free benchmark runs for any LEZ team needing proving cost numbers.
I saw the lssa-zkvm-testing repo from the Nescience team and the October 2024 note about planning CUDA benchmarks for RISC Zero and SP1. Just left a comment there with independent GPU results and posting here so builders are aware this infrastructure exists.
What I can give you:
-
Cycle count per guest program
-
Composite STARK + succinct compress timing
-
Total proving time
-
GPU vs CPU speedup comparison
-
Proof size
-
5-run variance (min/max/mean/σ)
Recent example: NIST PQ signatures in RISC Zero:
SLH-DSA-128s (SHA-256-heavy, similar profile to keccak256-heavy ECDSA):
-
CPU baseline: 1,731s (28m 51s)
-
RTX 4090 GPU mean: 16.33s (σ ±0.55s)
-
Speedup: 106x
Full 5-run variance data + proof artifacts: https://github.com/NthMOMENT/ghostprover-benchmarks
Why this matters for LEZ: Every LEZ program compiles to a RISC Zero guest binary. Proving cost directly affects your architecture decisions — heartbeat budget, push vs pull, whether a native precompile follow-on makes sense. Getting real GPU numbers before committing saves weeks.
If you have a guest ELF or reproducible build, send it my way. I’ll run it on GPU, publish results publicly, and you can cite the numbers in your proposal or M0 deliverable.
Free for cost baselines. No strings attached.
GhostProver / NTH MOMENT
https://github.com/NthMOMENT/ghostprover-benchmarks