AMD released ZenDNN 6.1, its oneDNN API-compatible deep learning library, Phoronix reports. AMD is still pushing ZenDNN even as it works with Intel on PyTorch/oneDNN optimizations.
- More work on the LowOHA inference stack
- Faster group MATMUL kernels, FP16 and DA8W4 kernels
- MATMUL kernels for DA8W4, DA8W8, INT8 and BF16, plus symmetric INT8 prepacking
My take
CPU inference keeps getting better on Zen, which is good news for anyone running small models without a big GPU.
God Bless, and Happy hacking!
