Jonathan Bentz comments

Results 8 comments of


                                            Jonathan Bentz

nvcc fatal: Option '--ptx (-ptx)' is not allowed when compiling for multiple GPU architectures

As of CUDA 12.8 we have updated the CUDA samples build system to CMAKE. Closing this issue for now, and if you still have problems with the current v12.8 Samples,...

Copy engine vs SM mode

The CUDA Samples are intended for illustration and functional use, but are NOT intended for serious performance testing. We'd recommend you use the [NVBandwidth](https://github.com/NVIDIA/nvbandwidth) code to do performance testing.

P2P is supported between GPUs that belong to different NUMA nodes

Do you mind trying the following commands to see what the driver says is supported? `$ nvidia-smi topo -p2p r` `$ nvidia-smi topo -p2p w`

cuda_driver_api.h doesn't exist?

Which version of the CUDA Samples are you referring to, and where is `cuda_driver_api.h` referenced? My search in v12.8 Samples doesn't show any use of `cuda_driver_api.h`

p2pBandwidthLatencyTest has an oom error

For this kind of bug it will be difficult for us to triage remotely. Recommend you join our Developer Program and then file a bug on our website here. https://developer.nvidia.com/bugs/new

Run busGrind -p 1 -u 0 -e 0 -d 1, the Concurrent Host/Device Bandwidth Matrix result is confusing.

For this kind of question we will refer you to ask your question in the NVIDIA Developer Forums https://forums.developer.nvidia.com/ . Thanks.

Is there any sgemm example ( e.g. fp32) ?

https://github.com/NVIDIA/cuda-samples/blob/master/Samples/4_CUDA_Libraries/simpleCUBLAS/simpleCUBLAS.cpp shows the use of CUBLAS sgemm.

Does the cuda sample program have a bandwidth test program with TMA?

Please keep in mind that the Samples are intended for illustration only. They are not designed to be performance tests and shouldn't be used as such. Thanks.