Intel AMX Support Continues Being Prepped For The Linux Kernel
Intel engineers have been publishing open-source/Linux enablement patches around Advanced Matrix Extensions (AMX) for nearly one year now. While Intel Xeon "Sapphire Rapids" with AMX support is expected around the end of this year, one of the key pieces yet to land is the Linux kernel support.
Sent out on Sunday was the fifth iteration of the kernel patches for supporting Intel Advanced Matrix Extensions. AMX is Intel's new programming paradigm with a focus on better AI performance both for training and inference. AMX is built around the concept of "tiles" as a set of two-dimensional registers for representing a larger memory image and accelerators that can operate on said tiles. The initial AMX implementation is focused on BFloat16, TILE, and INT8 while the design is extensible for expanding in the future.
Intel has sent out and since merged the AMX toolchain support while the Linux kernel bits are what have yet to land. The kernel code includes wiring up the eXtended Feature Disabling (XFD), exception handler changes, and related bits for the dynamic user-state management and getting AMX working along with initial optimizations.
With the v5 patches sent out on Sunday the code has been re-based to the Linux 5.13 Git state, some simplifications to the code made, and a variety of other low-level code improvements for this AMX handling.
The latest AMX kernel patches can be found on the LKML.
We'll see if this work gets wrapped up in time for merging into the Linux 5.14 cycle kicking off around the start of July. If the AMX kernel portion doesn't land for Linux 5.14, it begins to risk not seeing the functionality supported by the stock kernels of H2'2021 Linux distribution updates which in turn could impact the AMX exposure at launch -- assuming Sapphire Rapids is still tracking for shipping before the end of this calendar year. Normally Intel gets their kernel changes for new hardware features mainline well in advance of launch but unless Sapphire Rapids is running late this could be one of the times where the support is merged rather close to launch
Sent out on Sunday was the fifth iteration of the kernel patches for supporting Intel Advanced Matrix Extensions. AMX is Intel's new programming paradigm with a focus on better AI performance both for training and inference. AMX is built around the concept of "tiles" as a set of two-dimensional registers for representing a larger memory image and accelerators that can operate on said tiles. The initial AMX implementation is focused on BFloat16, TILE, and INT8 while the design is extensible for expanding in the future.
Intel has sent out and since merged the AMX toolchain support while the Linux kernel bits are what have yet to land. The kernel code includes wiring up the eXtended Feature Disabling (XFD), exception handler changes, and related bits for the dynamic user-state management and getting AMX working along with initial optimizations.
With the v5 patches sent out on Sunday the code has been re-based to the Linux 5.13 Git state, some simplifications to the code made, and a variety of other low-level code improvements for this AMX handling.
The latest AMX kernel patches can be found on the LKML.
We'll see if this work gets wrapped up in time for merging into the Linux 5.14 cycle kicking off around the start of July. If the AMX kernel portion doesn't land for Linux 5.14, it begins to risk not seeing the functionality supported by the stock kernels of H2'2021 Linux distribution updates which in turn could impact the AMX exposure at launch -- assuming Sapphire Rapids is still tracking for shipping before the end of this calendar year. Normally Intel gets their kernel changes for new hardware features mainline well in advance of launch but unless Sapphire Rapids is running late this could be one of the times where the support is merged rather close to launch
2 Comments