Intel NPU Driver Being Refactored For More Versatile CPU+NPU Handling

Intel's open-source NPU "iVPU" Linux kernel driver for supporting their Neural Processing Unit beginning with Core Ultra "Meteor Lake" processors is already seeing a code refactoring. The refactoring of this Intel accelerator driver is intended for allowing more versatile CPU and NPU combinations moving forward.

Intel recently posted a set of patches to refactor the hardware layer with their NPU (iVPU) driver. The patch cover letter explains how they are better separating the NPU buttress and NPU IP components for handling different configurations in future CPU models:

"The NPU device consists of two parts: NPU buttress and NPU IP. Buttress is a platform specific part that integrates the NPU IP with the CPU. NPU IP is the platform agnostic part that does the inference.

This refactor enables support for multiple platforms using a single NPU IP, so for example NPU IP 37XX could be integrated into MTL and LNL platforms."

Thus better handling if the NPU IP is reused by future/multiple hardware generations and other combinations that could come about with Intel ramping up their NPU efforts as part of their overall AI ambitions.

This patch series is currently restructuring around 2.7k lines of code within the Intel NPU accelerator driver. It will be interesting to see what the future holds on the Intel NPU front. At least Intel already has their open-source driver support in place for Meteor Lake as well as early work for Arrow Lake and Lunar Lake. Meanwhile earlier this year brought an open-source AMD "XDNA" driver for Ryzen AI on Linux but we haven't seen too much activity there nor any communications around trying to upstream that kernel driver or what all they have as far as Ryzen AI support plans on Linux.