Test failures on Nvidia GH200

[giordano]

Some test fails on Nvidia GH200, according to Tim these are known issues due to new dynamic parallelism API , but I'm opening this ticket to track them.

Here is the full output of Pkg.test("CUDA") on this GPU: cuda-tests-gh200.log

Main issue, repeated several times, is

  CUDA error: limit is not supported on this architecture (code 215, ERROR_UNSUPPORTED_LIMIT)

[maleadt]

Dup of #1846? I don't see new errors in the log.

[giordano]

Ah, I searched for "GH200" in the repo but didn't find anything before opening the issue 😛

Not mentioned in that page I see

  Failed to compile PTX code (ptxas exited with code 255)
  Invocation arguments: --generate-line-info --compile-only --verbose --gpu-name sm_90 --output-file /tmp/jl_HKnQTqFMVs.cubin /tmp/jl_hEHiJKzsJy.ptx
  ptxas /tmp/jl_hEHiJKzsJy.ptx, line 291; fatal   : Unsupported Function 'cudaDeviceSynchronize' on arch 'sm_90' or higher
  ptxas fatal   : Ptx assembly aborted due to errors

which I guess is related to #2052