Compute Accelerator Subsystem Being Introduced For Linux 6.2

[phoronix]

Phoronix: Compute Accelerator Subsystem Being Introduced For Linux 6.2

It's happening: the new "accel" compute accelerator subsystem is now queued for introduction with the Linux 6.2 kernel once that merge window opens in December...

https://www.phoronix.com/news/Linux-6.2-Compute-Next

[unic0rn]

My first reaction:

image.png

Sorry, couldn't resist.

[coder]

Some questions, in case anyone knows:

1. For GPUs with substantial deep learning capabilities, are they expected to use a different driver for exposing that functionality? If yes, how would that driver interact with the graphics driver/functionality?
2. Does the new subsystem have a requirement for open source userspace software that can utilize the device?
3. Does the driver API address cache-coherent device memory (e.g. via CXL), in any explicit way, or have any other functional advantages over using the GPU driver subsystem?