cuptiProfilerGetCounterAvailability failed on sm_90

Developer Tools CUPTI – CUDA Profiler Tools Interface
1

Hi NVIDIA folks!

I successfully run the official sample cupti_metric_properties on v100 platform, which reveals all available metrics on this GPU.

Yet I failed to run this sample on h100 platform. This root cause is:

CUPTI need to call cuptiProfilerGetCounterAvailability twice to obtain the counter availability image. The first call is to obtain the size of the image, then the user sets the buffer as corresponding size, finally the second call returns the actual image.

The sample failed at the second call to cuptiProfilerGetCounterAvailability, the return value is 999 (give nothing valuable).

My environment is :

  • CUDA 12.1
  • Ubuntu 20.04
  • Driver 555.42.06
  • NVIDIA H100

Could you please help to solve this issue? Thanks so much

2

Hi, @Zobin

Sorry for the issue you met. There may be CUDA/driver not compatible issue here.
Would you please use 12.1+R530 driver or 12.5+R555 instead ?

3

Thanks veraj, after changing to 12.5 + R555, this issue solved.

4

Good to know. Thanks !

5

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.