Intel Meteor Lake's VPU Linux Driver Updated, UMD Code Posted
Back in July Intel engineers published the initial open-source driver code around the new Versatile Processing Unit "VPU" coming with Meteor Lake. This VPU block with 14th Gen Core CPUs is intended for AI inference acceleration for deep learning software.
Since then the new Intel VPU driver has continued to be revised and has been adapted to the Linux kernel's new "accel" compute accelerator subsystem. Out today is the fifth version of their kernel driver as well as having recently posted their user-space driver code for the VPU.
Out this morning is the iVPU v5 patches for this AI inference accelerator. The revised patches have some optimizations, continue renaming code in being adapted from the DRM subsystem to the new "accel" framework/subsystem, and other clean-ups. Overall though things are settling down on this driver and it looks like in the next kernel cycle or two we could see this iVPU accelerator driver merged to the mainline kernel.
In December Intel also quietly posted the "linux-vpu-driver" as the user-space driver for interfacing with the VPU kernel driver. This user-mode driver in turn pushes workloads to the kernel driver and implements the oneAPI Level Zero interface. It's through oneAPI Level Zero or building off software like Intel OpenVINO that deep learning software will be able to leverage the Versatile Processing Unit when Meteor Lake ships.
The kernel patches continue to indicate though that the VPU will require a "closed source binary" as firmware, similar to GPUs. So for those hoping it would have open-source firmware for the VPU, that doesn't appear to be the case at this time.
Since then the new Intel VPU driver has continued to be revised and has been adapted to the Linux kernel's new "accel" compute accelerator subsystem. Out today is the fifth version of their kernel driver as well as having recently posted their user-space driver code for the VPU.
Out this morning is the iVPU v5 patches for this AI inference accelerator. The revised patches have some optimizations, continue renaming code in being adapted from the DRM subsystem to the new "accel" framework/subsystem, and other clean-ups. Overall though things are settling down on this driver and it looks like in the next kernel cycle or two we could see this iVPU accelerator driver merged to the mainline kernel.
In December Intel also quietly posted the "linux-vpu-driver" as the user-space driver for interfacing with the VPU kernel driver. This user-mode driver in turn pushes workloads to the kernel driver and implements the oneAPI Level Zero interface. It's through oneAPI Level Zero or building off software like Intel OpenVINO that deep learning software will be able to leverage the Versatile Processing Unit when Meteor Lake ships.
The kernel patches continue to indicate though that the VPU will require a "closed source binary" as firmware, similar to GPUs. So for those hoping it would have open-source firmware for the VPU, that doesn't appear to be the case at this time.
2 Comments