cudaMemcpyPeerAsync behavior for different hardware

[Robert_Crovella]

@samuel.gozel

the last test (latency with P2P enabled) was running for over 15minutes,

That’s abnormal.

@samuel.gozel

so I reduced the repeat to 10 instead of 10k

That should not be necessary. The test is designed to run in less than a minute, as-is, on a proper modern platform (try it unmodifed on your DGXA100 system if you like.)

@samuel.gozel

Unidirectional P2P=Enabled Bandwidth Matrix (GB/s)
   D\D     0      1      2      3      4      5      6 
     0 321.04   *0.73   *0.73  18.39  18.46  18.47  18.48 

The results above marked with an asterisk, for example, are a problem/unexpected. I think the platform is suspect. If you ordered it configured exactly like this from a reputable OEM, then you should take these results to them and ask for resolution. That isn’t something we can do here, nor can NVIDIA fix your platform.

If this is a platform you built yourself, then I would suspect that you have started with a motherboard setup which was not properly designed for this activity.

I don’t think I will be able to help further with a platform issue, but this topic may be of interest.