Intel NPU Acceleration Library 1.1 Adds New Model Support & Optimizations

[phoronix]

Phoronix: Intel NPU Acceleration Library 1.1 Adds New Model Support & Optimizations

Intel software engineers did a late Friday night release of the Intel NPU Acceleration Library v1.1, their Python library for tapping into the Intel Neural Processing Unit (NPU) found on the new Core Ultra (Meteor Lake) processors. This Python library makes it easier to interface with the Intel VPU/NPU kernel driver and in turn enjoying accelerated operations for AI...

https://www.phoronix.com/news/Intel-NPU-Accel-Lib-1.1

[oleid]

The big question is who is going to use this on Linux. One would need a way that would free potential users to implement support for specific hardware; we need some kind of abstraction layer for ML stuff supporting different vendors, i.e. AMD and Intel TPUs as well as NVidia, AMD and Intel GPUs. Microsoft, for example, provides DirectML on windows.

[luno]

The big question is who is going to use this on Linux. One would need a way that would free potential users to implement support for specific hardware; we need some kind of abstraction layer for ML stuff supporting different vendors, i.e. AMD and Intel TPUs as well as NVidia, AMD and Intel GPUs. Microsoft, for example, provides DirectML on windows.

I used https://github.com/openvinotoolkit/openvino as directml alternative in native webnn, it works pretty well. Apple also has nice abstraction for doing ML, maybe someday freedesktop will create something for Linux