AMD XDNA Linux Driver v3 Published For Ryzen AI Upstreaming

[phoronix]

Phoronix: AMD XDNA Linux Driver v3 Published For Ryzen AI Upstreaming

AMD engineers continue work toward upstreaming their XDNA kernel driver for Linux in enabling the Ryzen AI NPU on open-source. The "v3" patches were posted on Wednesday but given the timing it looks like it will be missing out still on merging for the upcoming Linux 6.12 LTS cycle...

https://www.phoronix.com/news/AMD-XDNA-Linux-Driver-v3

[chuckula]

Good to see more open sourcing for these accelerators and I'm expecting a DRI-like standardization to build up cross-vendor AI accelerator infrastructure.

[lejeczek]

Now it's an ageing story, sad story - https://www.youtube.com/watch?v=U1SC0kfXlnw - one of many.

[Spacefish]

Now it's an ageing story, sad story - https://www.youtube.com/watch?v=U1SC0kfXlnw - one of many.

Nah, that´s ok.. The demo was kind of funny though.. I feel kind of followed b.t.w. the guy in the YT video mentions my PR #1

The "XDNA" NPU, is just a Versal AI core from Xilinx configured to only support some basic datatypes like int8.. The interface is almost identical to the existing linux drivers even back then, they just did not release the changes that fast / made them available.

The demo was just a python script, using a YOLOv5 net with a basic classifier to classify some images running in the ONNX runtime, you could easily just run this on the CPU or any GPU / accelerator supported by onnx.
It really looked like it was hacked together within 2-3 days, they even left in the auto generated electron-quick-start folder ^^

The NPUs are not really that interesting in this generation, a mid-range GPU is faster for inference and can do training, the NPU can´t be really used for training as it does not support floating point datatypes in high enough precision..

These NPUs will be mainly used to optimize power usage on mobile systems, so you can blur your background in Teams and do Noise Filtering in the Audio stream without draining your battery that fast... The "Recall" feature could easily run on CPU or the GPU.

ROCm on consumer GPUs works now back to RDNA1 (RDNA 2+ fully works RDNA1 limited)... But you still have to build it from source if you want the best performance.. The RDNA3 GPUs are essentially all supported without any issue, you just have to set one environment variable.

I am more interested in the upcoming "UDNA" where they will join the datacenter GPUs ("CDNA") and Consumer RDNA GPU architectures again. Hopefully we will see a consumer card with HBM memory again, like the Vega64 that would be great

At least for large models, the memory bandwidth is the bottleneck currently!

[zamroni111]

i tried to compile in ubuntu 22 with gcc12 but compile error.
you can try by cloning this repo:

https://github.com/intel-lab-lkp/linux/tree/Lizhi-Hou/accel-amdxdna-Add-documentation-for-AMD-NPU-accelerator-driver/20240912-020807