Intel Compute Runtime 24.39.31294.12 Fixes Lunar Lake OpenCL, Disables Ice Lake & Older

[Svyatko]

I really hope we see AMD's ROCm and Intel's Compute Runtime eventually give us that kind of longevity from RDNA 1+ and Xe+.

RDNA1 has no official support for GPGPU and it will not get it in future.
You can use AMD's OpenCL PAL driver or Rusticl to get OpenCL support.

[Svyatko]

So just Xe / Gen 12 and newer are officially supported. That sucks. I do wonder if the pre-Xe iGPUs are performant enough to be useful compared to their CPU cores. But at a minimum, being able to easily play with their compute runtime on almost all their consumer CPUs from the last 10 years seems like it would add value for the ecosystem (like the ubiquity of CUDA support does on the NVIDIA side). I wonder if distros will flip the switch to enable the now deprecated generations.

Pre-Xe iGPUs are too slow for GPGPU. Possibly Intel wants to tie GPGPU software with Xe kernel driver.
Rusticl provides OpenCL 3.0 support for older and newer Intel GPUs.

[varikonniemi]

intel is good, as long as you buy new and move on within warranty time. Astonishingly bad support.

[geerge]

Meanwhile in NVIDIA land, their cards are generally a pain in the ass in Linux, but the CUDA support is exemplary.

• On Windows, the now > 10 year old Kepler cards have a WHQL driver for Windows 11 and will continue working for years to come including GPU compute
• On Linux, the earliest Maxwell cards are now > 10 years old (e.g. Quadro K2200) and are still supported by the current R550 production branch

I really hope we see AMD's ROCm and Intel's Compute Runtime eventually give us that kind of longevity from RDNA 1+ and Xe+.

AMD has a bad habit of ditching support in rocm way too quick, and supporting new architectures in rocm way too slow. Very bad combo. Now that they're in a good position money-wise they need to hire dozens of devs permanently to get their software shit in order. Not this pissy one or two hires every now and then that phoronix covers and people get excited over.

[pWe00Iri3e7Z9lHOX2Qx]

Meanwhile in NVIDIA land, their cards are generally a pain in the ass in Linux, but the CUDA support is exemplary.

• On Windows, the now > 10 year old Kepler cards have a WHQL driver for Windows 11 and will continue working for years to come including GPU compute
• On Linux, the earliest Maxwell cards are now > 10 years old (e.g. Quadro K2200) and are still supported by the current R550 production branch

I really hope we see AMD's ROCm and Intel's Compute Runtime eventually give us that kind of longevity from RDNA 1+ and Xe+.

[nyanmisaka]

nyanmisaka Thanks for finding that out. So we are back to facing Ice Lake and Comet Lake stuff, which are mostly first launched in 2019. Intel launched Comet Lake Refresh in 2021. Oh well... hopefully nobody bought that for GPU computation.

Ice Lake Gen11 Gfx only appears in laptops and some mini PCs, while Comet Lake uses the same Gen9 Gfx as Sky Lake launched in 2015. AMD's last generation of GCN was in the same boat, with Radeon VII released in 2019 but discontinued ROCm and Windows latest driver support last year. Sad to see this.

[pWe00Iri3e7Z9lHOX2Qx]

So just Xe / Gen 12 and newer are officially supported. That sucks. I do wonder if the pre-Xe iGPUs are performant enough to be useful compared to their CPU cores. But at a minimum, being able to easily play with their compute runtime on almost all their consumer CPUs from the last 10 years seems like it would add value for the ecosystem (like the ubiquity of CUDA support does on the NVIDIA side). I wonder if distros will flip the switch to enable the now deprecated generations.

[billyswong]

nyanmisaka Thanks for finding that out. So we are back to facing Ice Lake and Comet Lake stuff, which are mostly first launched in 2019. Intel launched Comet Lake Refresh in 2021. Oh well... hopefully nobody bought that for GPU computation.

[nyanmisaka]

The "legacy" rocket lake processor was launched in 2021. Wow.

It looks like it's a typo, Rocket Lake (RKL) uses Gen12 Gfx like Tiger Lake (TGL), and has not been removed from the repo like Gen11-

https://github.com/intel/compute-runtime/blob/605a1ae0abc988c764895c750d539269ff499d30/cmake/setup_platform_flags.cmake#L11

edit: confirmed it's a typo. https://github.com/intel/compute-run...ent-2429067721

[pioto]

They should move on with their default llvm backend, as currently their compute runtime is removed from Debian Testing, due to the dependency on retired llvm.