LLVM Clang Begins Picking Up AMX-FP16 Support For Granite Rapids

[phoronix]

Phoronix: LLVM Clang Begins Picking Up AMX-FP16 Support For Granite Rapids

In addition to GCC landing patches this week in preparing for Grand Ridge and Sierra Forest, the LLVM/Clang open-source compiler stack has also been seeing patches to prepare for future Intel server processors...

https://www.phoronix.com/news/LLVM-Clang-AMX-FP16

[ms178]

Intel engineers might have been a bit too eager with that feature though, as it seems they enabled it even for older CPUs that don't support it, too. At least I got a ton of '-amxfp16' is not a recognized feature for this target (ignoring feature) since that change and filed this bug report to make them aware of it.

[brad0]

Intel engineers might have been a bit too eager with that feature though, as it seems they enabled it even for older CPUs that don't support it, too. At least I got a ton of '-amxfp16' is not a recognized feature for this target (ignoring feature) since that change and filed this bug report to make them aware of it.

Yes, typos do happen. Thanks for noticing.

[coder]

It's intriguing to see this is fp16, and not bf16. I could see it being useful for classical image processing (esp. HDR) and spatial/depth data.

[jayN]

Interesting to see the latest MLID comments on Granite Rapids ... suggesting it is being moved up to 1H 2024. This would be strange, following so closely behind Emerald Rapids, but it will also use a different socket ... so maybe the thinking is that EMR will be for SPR refresh while Granite Rapids will be for new sockets.

[jayN]

I'd expect FP8 support in AMX eventually, since there was already announced add-on support for Gaudi2.

[coder]

Interesting to see the latest MLID comments on Granite Rapids ... suggesting it is being moved up to 1H 2024. This would be strange, following so closely behind Emerald Rapids, but it will also use a different socket ... so maybe the thinking is that EMR will be for SPR refresh while Granite Rapids will be for new sockets.

Yeah, that's a little weird, if true. I wonder if Emerald Rapids is being kept alive as a fallback, in case there turn out to be delays in Granite Rapids.

According to the leaks, core counts are likely to skyrocket. I find that a little far-fetched, but we'll see. Also, they show PCIe 5.0, whereas I'd have expected them already to be moving on to PCIe 6.0.