[test] amd-demo: MI350 optimized kernels benchmark #858
+1,045,030
−12
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
sunxxuns
requested review from
Michael-Equi,
jimmyt857,
kvablack and
uzhilinsky
as code owners
sunxxuns
changed the title
- Aiter Flash Attention for AMD GPUs - Triton kernels for RMSNorm, GELU+Mul - Full policy inference: 142ms latency, 7Hz (Pi0 3.5B, batch=1) - 8-GPU DDP training: 407 samples/s (3.3B model) - Training convergence verified - Perfetto traces included
sunxxuns
pushed a commit
to sunxxuns/openpi-1
that referenced
this pull request
Feb 2, 2026
Benchmark results comparing NVIDIA H200 vs AMD MI350 (PR Physical-Intelligence#858): Inference (Pi0 3.5B, batch=1): - H200: 118.5-120.8 ms (8.28-8.44 Hz) - MI350: 142.0 ms (7.04 Hz) - H200 is ~15-17% faster 8-GPU DDP Training (3.3B model): - H200 (SDPA): 320 samples/s best - MI350 (Aiter): 407 samples/s best - MI350 is ~21-27% faster Files added: - BENCHMARK_H200.md: Full comparison results - scripts/benchmark_h200_ddp.py: DDP training benchmark (eager) - scripts/benchmark_h200_ddp_sdpa.py: DDP training benchmark (SDPA) Trace files generated but not committed (too large): - h200_inference.json (280 MB) - h200_ddp_*_rank[0-7].json (~41-46 MB each)
sunxxuns
pushed a commit
to sunxxuns/openpi-1
that referenced
this pull request
Feb 9, 2026
Benchmark results comparing NVIDIA H200 vs AMD MI350 (PR Physical-Intelligence#858): Inference (Pi0 3.5B, batch=1): - H200: 118.5-120.8 ms (8.28-8.44 Hz) - MI350: 142.0 ms (7.04 Hz) - H200 is ~15-17% faster 8-GPU DDP Training (3.3B model): - H200 (SDPA): 320 samples/s best - MI350 (Aiter): 407 samples/s best - MI350 is ~21-27% faster Files added: - BENCHMARK_H200.md: Full comparison results - scripts/benchmark_h200_ddp.py: DDP training benchmark (eager) - scripts/benchmark_h200_ddp_sdpa.py: DDP training benchmark (SDPA) Trace files generated but not committed (too large): - h200_inference.json (280 MB) - h200_ddp_*_rank[0-7].json (~41-46 MB each)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.