Investigate Performance of `deterministic DeviceReduce (RFA)` on `B200`
- Dominant language
- C++
- Stars
- 2.5k
- Forks
- 486
- Avg merge
- 2d 6h
- Merged PRs (30d)
- 295
Description
Recent benchmark's on **B200** of **deterministic DeviceReduce** i.e with `gpu_to_gpu` **determinism** showed that the performance is only 45% and 51% of SOL with F32 and F64 types respectively. But, the performance of **default DeviceReduce** i.e `run_to_run` **determinism** is about 80% and 86% percent with F32 and F64 types.
Ideally we expect the `gpu_to_gpu` determinism performance to be closer to `run_to_run` determinism.
But the gap is too huge, hence performance investigations are needed.
Tuning **deterministic DeviceReduce** with only two tunings PARAMs `BLOCK_THREADS` and `ITEMS_PER_THREAD` did not reveal to any improvements. So, we should consider adding more tuning parameters such as `BLOCK_ALGORITHM`, `ITEMS_PER_VEC_LOAD` etc.. to find better tunings.
Contributor guide
Assessment
This issue has not been assessed yet.