Hi
Are there any plans to support intel arc hardware accelerated Xe Matrix Engines (XMX) blocks on the arc cards?
@MPdaedalus depends what it would take. If it requires writing new c++ code that might be some extra work. Maybe if onednn supports it we can add support for certain architectures but otherwise I have to do more research. Feel free to file an issue.
Edit: Just looking at it, does it even have any market share? Everything I’m reading is more just assuring people the card is alive than seeing any benchmarks for it.
its no big deal if you won’t support it, was just wondering as I was reading about them recently and interested how they would fare compared to Tensor cores and the fact they offer 16GB for a reasonable price.
There is not much info about them for ML but there is this article if you want to go into them a bit>
@MPdaedalus sure thanks! I just haven’t seen much on it and not much interest so far. I’m more inclined to do ROCM/AMD focused stuff but this doesn’t look bad to do at some point if the chip gains marketshare.
could not agree more, ROCM support would be great and allow DL4J to be more widely distributed on end user platforms. Anything that helps break the NVIDIA ML monopoly.