GPU correctness board
Every GPU here ran sdcguard check: a real transformer workload with every matrix multiply
checked against checksums, a memory pattern test, speed tests, and the same fixed model run three times to see if its
output stays bit-identical. Updated 2026-10-28 13:05 UTC.
By GPU model
| GPU | runs | distinct GPUs | PASS | FAIL | SUSPECT | median matmul TFLOPS | of datasheet peak | median GB/s | slowed down | same fingerprint |
|---|---|---|---|---|---|---|---|---|---|---|
| NVIDIA L4 | 3 | 3 | 3 | 0 | 0 | 55 | 46% | 231 | 0 of 3 | 3 of 3 |
| NVIDIA H100 80GB HBM3 | 2 | 2 | 2 | 0 | 0 | 685 | 69% | 3,033 | 0 of 2 | 2 of 2 |
| NVIDIA A100 80GB PCIe | 1 | 1 | 1 | 0 | 0 | 230 | 74% | 1,670 | 0 of 1 | 1 of 1 |
| NVIDIA B200 | 1 | 1 | 1 | 0 | 0 | 1,326 | 59% | 6,415 | 0 of 1 | 1 of 1 |
| NVIDIA H200 | 1 | 1 | 1 | 0 | 0 | 670 | 68% | 4,240 | 0 of 1 | 1 of 1 |
| NVIDIA L40S | 1 | 1 | 1 | 0 | 0 | 173 | 48% | 644 | 0 of 1 | 1 of 1 |
| Tesla T4 | 1 | 1 | 1 | 0 | 0 | 23 | 36% | 233 | 0 of 1 | 1 of 1 |
"Same fingerprint": runs of the identical setup (GPU model, software versions) whose fixed-model output matched the most common one. A healthy GPU should always match.
By provider (as reported)
| provider | runs | GPU models | PASS | FAIL | SUSPECT | median of datasheet peak | slowed down |
|---|---|---|---|---|---|---|---|
| Modal | 10 | NVIDIA A100 80GB PCIe, NVIDIA B200, NVIDIA H100 80GB HBM3, NVIDIA H200, NVIDIA L4, NVIDIA L40S, Tesla T4 | 10 | 0 | 0 | 59% | 0 of 10 |
Recent results
| date | result | GPU | provider | TFLOPS | results checked | fingerprint | source |
|---|---|---|---|---|---|---|---|
| 2026-10-18 | PASS | NVIDIA A100 80GB PCIe | Modal | 230 | 101,834 | a10587e20440 | modal_a100_80gb_c770d238.json |
| 2026-10-18 | PASS | NVIDIA B200 | Modal | 1,326 | 192,805 | c27adbfea641 | modal_b200_a203014e.json |
| 2026-10-18 | PASS | NVIDIA H100 80GB HBM3 | Modal | 685 | 241,583 | 5613afa2605e | modal_h100_a_45b2fbe6.json |
| 2026-10-18 | PASS | NVIDIA H100 80GB HBM3 | Modal | 683 | 243,711 | 5613afa2605e | modal_h100_b_9e83ee71.json |
| 2026-10-18 | PASS | NVIDIA H200 | Modal | 670 | 268,032 | 5613afa2605e | modal_h200_5c7bd569.json |
| 2026-10-18 | PASS | NVIDIA L40S | Modal | 173 | 302,747 | a10587e20440 | modal_l40s_713a5503.json |
| 2026-10-18 | PASS | NVIDIA L4 | Modal | 57 | 97,573 | a10587e20440 | modal_l4_a_c97e475e.json |
| 2026-10-18 | PASS | NVIDIA L4 | Modal | 55 | 84,917 | a10587e20440 | modal_l4_pair_d9cc6760.json |
| 2026-10-18 | PASS | NVIDIA L4 | Modal | 54 | 84,531 | a10587e20440 | modal_l4_pair_e1bca08b.json |
| 2026-10-18 | PASS | Tesla T4 | Modal | 23 | 65,630 | f70d4a7ce612 | modal_t4_933c7202.json |
Add your GPU: pip install sdcguard && sdcguard check --share, then press
Submit on the issue it opens. Results are self-reported and come from GitHub issues; each card carries a digest of its contents,
hardware IDs are hashed, and every FAIL or SUSPECT links to its issue. Simulated-fault demos, CPU runs and runs under a
minute are left out. Raw data: results.json, reference.json. Code and docs on GitHub.