On Friday AMD sent out another batch of AMDGPU and AMDKFD kernel driver feature patches destined for the upcoming Linux 6.15 kernel cycle. One notable feature in this late pull request is introducing a new “high precision” mode to be found with the GFX950 target, which is believed to be the upcoming Instinct MI350X series.
In prior weeks there have been AMDGPU/AMDKFD pull requests preparing more graphics driver features and new GPU support for introduction in Linux 6.15. Friday’s latest pull request was mostly about landing a number of fixes but there is also some minor feature work.
This newest pull brings VCN 5.0.1 IP updates, DCN 4.0.1 display fixes, SMU 14 fixes, SR-IOV fixes, RDNA4 / GFX12 DCC fixes, and even a HyperZ fix for old ATI RS400 graphics.
New on the feature side though is a new per-process flag for setting desired precision. The AMDKFD kernel compute driver sees the ability added to handle per-process flags and is followed by the ability for user-space to set desired precision beginning with GFX950.
GFX950 is the next iteration of AMD’s CDNA compute architecture. Given the timing and all things considered, GFX950 is believed to be the Instinct MI350X series. The Instinct MI350 series is expected to launch this calendar year and there’s been a lot of open-source Linux driver activity continuing around the GFX950 CDNA target.
What this new precision mode handling is with Linux 6.15 allows setting a new “HIGH_PRECISION” mode for the Matrix Fused Multiply Add (MFMA) instructions on the AMD matrix cores. With Linux 6.15+ and when also running the next versions of the ROCm compute stack, ROCm will pass on the new “HSA_HIGH_PRECISION_MODE” environment variable when set for enabling the high precision math mode. This MFMA high precision mode is only implemented with GFX950 and the HSA_HIGH_PRECISION_MODE control having no impact for other GPUs.
The AMDKFD kernel driver patches and ROCm patches don’t shed any further light on this high precision MFMA mode with upcoming GFX950 / Instinct MI350X hardware. AMD Matrix Cores already support FP64 / FP32 / FP16 / BF16 / INT8 data formats.
This latest set of AMDGPU/AMDKFD feature patches intended for the upcoming Linux 6.15 merge window can be found via this pull request.
