how can the L2 arithmetic intensity be less than the HBM AI.

[etiennemlb]

Describe your question

I have the following roofline:

This gives me a lower L2 ai than HBM ai. As all loads/stores that go through L2 should go through HBM we should have AI L1 >= AI L2 >= AI HBM right ?

Additional context

No response

[etiennemlb]

Additionally, how would one interpret the AI L1/L2 dots. Would that mean the farther left we are, the better the caches are used (we do more loads in caches/LDS). If the L1 dot is farther than L2 this means we use the L1, if the L2 is farther than the HBM that means we use it ?

[ppanchad-amd]

Hi @etiennemlb. Internal ticket has been created to assist with your issue. Thanks!

[systems-assistant[bot]]

This issue has been migrated to: https://github.com/ROCm/rocm-systems/issues/37

[amd-hsivasun]

Imported to ROCm/rocm-systems