Run kernels whose machine code was written by Omega on your own NVIDIA GPU, check every output word against a CPU reference, and put your card on this board. You need an NVIDIA driver, Python 3 and git. No CUDA toolkit, no compiler, no pip install.
git clone https://github.com/aien-dev/omega-gpu-bench && cd omega-gpu-bench && ./run
It writes a receipt under receipts/. Open a pull request with it to
github.com/aien-dev/omega-gpu-bench; CI checks it and, once merged, it shows up here.
What runs where. The kernels are SASS assembled by Omega's asm2 from a
hand-written listing and wrapped into a cubin by cubin.om; CI rebuilds them from source and checks
the bytes. They reach your GPU through NVIDIA libcuda (cuModuleLoadData + cuLaunchKernel): an interop oracle for Omega's machine-code encoder and cubin writer, not CUDA++'s own user-space driver path. Receipts are self-reported: CI checks the kernel hashes,
counts, verdicts, every reported reference value (recomputed from the public seed), the output hash of a
clean run, and the GB/s arithmetic, but nobody re-ran your GPU. The three seed rows are my own rentals on
vast through OpenResearch on 2026-10-26, run before this kit existed; they were not timed.
| # | GPU | Driver | Where | Origin | Kernels | vecadd speed (226 words) | Date | Receipt |
|---|---|---|---|---|---|---|---|---|
| 1 | NVIDIA GeForce RTX 4090 sm_89 | 580.65.06 CUDA driver API 13000 | vast (rented through OpenResearch) | imported CloudKernels run (before the kit) | neg sm80 NEG_DETECTED neg sm89 NEG_DETECTED vecadd sm80 PASS vecadd sm89 PASS | not timed | 2026-10-26 | 24bc78ab006e |
| 2 | NVIDIA A100-PCIE-40GB sm_80 | 570.211.01 CUDA driver API 12080 | vast (rented through OpenResearch) | imported CloudKernels run (before the kit) | neg sm80 NEG_DETECTED vecadd sm80 PASS | not timed | 2026-10-26 | aec125c67f2d |
| 3 | NVIDIA GeForce RTX 3090 sm_86 | 580.159.03 CUDA driver API 13000 | vast (rented through OpenResearch) | imported CloudKernels run (before the kit) | neg sm80 NEG_DETECTED neg sm86 NEG_DETECTED vecadd sm80 PASS vecadd sm86 PASS | not timed | 2026-10-26 | 6b9a64a187d5 |
vecadd: c[i] = a[i] + b[i] on 220 + 7 words, every word compared. neg: the same kernel with one register changed, which must be caught (NEG_DETECTED). Speed: median of 5 rounds of 20 launches timed with CUDA events, 12 bytes moved per word. 3 receipts shown, 0 rejected by the validator.
Source, rules and validator: github.com/aien-dev/omega-gpu-bench. Machine-readable: receipts.json.