Benchmarks¶
Nightly snapshot
GPU NVIDIA H200 · commit 67250e8172fa · run date 2026-10-16 · 192 ops, 1988 workloads
· nightly run
Page rendered 2026-10-18 01:08 Beijing time (2026-10-17 17:08 UTC) from the latest snapshot.
Environment¶
| image | ghcr.io/tile-ai/tileops-runner:cu132-torch2.13-tl-0.1.12 |
| gpu | NVIDIA H200 |
| driver | 580.178.04 |
| cuda | 13.2 |
| torch | 2.13.0+cu132 |
| tilelang | 0.1.12 |
| timer | cupti |
| memory_clock_mhz | 3201.0 |
| mig | Disabled |
| power_limit_w | 700.0 |
| sm_clock_max_mhz | 1980.0 |
| sm_clock_mhz | 1500.0 |
| tileops | 0.0.1.dev1338+g67250e817 |
Method¶
- One process, common inputs. Every implementation of an op is timed on the same tensors in the same process, in forward and then reversed order so drift does not land on whichever ran last.
- A fixed warmup and measurement budget per implementation, reported as the median over however many samples fit in it, with L2 cleared between iterations.
- Compilation and workspace setup excluded.
- Device time is what is compared — the union of the intervals the device spent executing the call's kernels, collected through CUPTI. A run that cannot collect device activity fails rather than falling back to a different clock.
Coverage¶
- 192 of 192 ops are measured against a real alternative — a tuned library kernel or a native PyTorch op — on the identical workload. The rest run against an eager reference only, which is not a bar worth reporting a win against.
- Absent from every table: 0 workloads errored and 1 were skipped in this run.
Data¶
| Page | Ops | Workloads |
|---|---|---|
| Elementwise | 70 | 527 |
| RoPE | 5 | 23 |
| Reduction | 19 | 176 |
| Normalization | 10 | 112 |
| Conv & Pool | 16 | 120 |
| GEMM | 6 | 108 |
| Quantization & Dequantization | 10 | 71 |
| Attention | 15 | 223 |
| MoE | 12 | 132 |
| Sampling | 6 | 48 |
| Linear Attention | 10 | 385 |
| SSM | 7 | 36 |
| Other | 6 | 27 |