xpu-smi does not enumerate B580 or Raptor Lake GPUs, breaking many Intel published examples

[jketreno]

Heya!

I have been trying to get any of the recent Intel AI/ML examples to work with a new system build with a raptor lake + b580.

While I can get an Ubuntu Desktop (Oracular 24.10) up and running no problem (great job there by that team!) I am unable to use any of the AI/ML examples to use the B580 or integrated GPU due to many of the frameworks now using 'xpu-smi' to determine if an Intel GPU exists:

To illustrate xpu-smi not finding the device:

xpu-smi discovery -j

However, sycl-ls does find the device:

sycl-ls

Kernel version being used (uname -a):

An example of current code that is trying to use xpu-smi in order to enumerate devices:

https://github.com/intel/intel-extension-for-pytorch/blob/af7dabeab01f5e571686cfe6bf89f5616e78d128/scripts/collect_env.py#L211

Is there a tagged version of xpu-smi I can build that supports these devices in enumeration?

The alternative would seem to be to have every framework team switch to using sycl-ls to determine if an xpu exists on the system.

I appreciate any advice, fixes, details, questions, or follow ups :)

Thanks,
James

[eero-t]

If project is already using / depending on SYCL, using sycl-ls seems much better / logical way of checking for GPU support, than XPU-SMI.

Especially as XPUM officially supports (= is tested with) only Intel Data Center GPUs, i.e. Flex + PVC. Whereas SYCL / L0 (that XPUM depends on) AFAIK officially support all (newer) Intel GPUs.