[FEA] Better grid size for H100 GPU with SXM5

[zhipeng93]

Is your feature request related to a problem? Please describe. When doing gemm in Hopper, we need to decide the grid size based on problem size, cluster shape and the hopper architectures.

Currently, cutlass only considered GH100 GPU but not H100 GPU with SXM5 [1]. They have different architectures [2]:

• GH100 GPU: 8 GPCs, 72 TPCs (9 TPCs per GPC), 144 SMs
• H100 GPU with SXM5: 8 GPCs, 66 TPCs (uneven TPC for GPCs?), 132 SMs

In H100 GPU with SXM5, cutlass may fail to find the optimal grid size, leading to sub-optimal performance. For example, when doing gemm mnk=(4096, 4096, 4096) on H100 GPU with SXM5 and using cluster shape (4, 2, 1), the result grid size is (4, 28, 1). However, cublas used grid size (60, 2, 1), and cublas has a better performance by 17%.

cutlass settings:

using TileShape           = Shape<_128,_128,_64>; 
using ClusterShape        = Shape<_4,_2,_1>;
cutlass::gemm::KernelTmaWarpSpecializedCooperative

[1] https://github.com/NVIDIA/cutlass/blob/main/include/cutlass/gemm/kernel/tile_scheduler_params.h#L249-L259 [2] https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/

Describe the solution you'd like Compute grid size according to the GPU archi.

Describe alternatives you've considered

Additional context

[zhipeng93]

When I try to hack the cutlass code for better performance on H100 GPU with SXM5, I find that when using cluster shape (4, 2, 1) or (2, 2, 1), only 120 SMs can be used. That is, when I set grid size greater than 120, the waves per SM would be greater than 1.

For example,

• cluster shape (4, 2, 1) and grid size (64, 2, 1), waves per SM is 1.07. (128/120~1.07)
• cluster shape (2, 2, 1) and grid size (62, 2, 1), waves per SM is 1.03. (124/120~1.03)
• cluster shape (4, 2, 1) and grid size (60, 2, 1), waves per SM is 1. (120/120~1)

I also check this setting: cluster shape (1, 1, 1) and grid size (132, 1, 1), waves per SM is 1.

Any ideas on the number of SMs that can be used when using thread block cluster?

[zhipeng93]

cc @hwu36 @Junkai-Wu

[hwu36]

@ANIKET-SHIVAM

[zhipeng93]

@ANIKET-SHIVAM Hi, can you help to explain this? :)

[ANIKET-SHIVAM]

@zhipeng93 yes, your observation is correct and is a known issue that for cluster sizes >= 4 optimal grid size is not launched. We should be able to elevate that limitation with cudaOccupancyMaxActiveClusters API. Will try to upstream changes soon.

[zhipeng93]

@ANIKET-SHIVAM Thanks for the reply!

It seems that I can never really get to use more than 120 sms when cluster sizes >=4, even I hack the heuristic for computing the grid size.

For example,

cluster shape (4, 2, 1) and grid size (64, 2, 1), waves per SM is 1.07. (128/120~1.07)
cluster shape (2, 2, 1) and grid size (62, 2, 1), waves per SM is 1.03. (124/120~1.03)

In Hopper SXM5, I should be able to use 132 sms. Can you also help to explain this?

[zhipeng93]

cc @ANIKET-SHIVAM @hwu36