Intel compiler engineer Feng Zou has upstreamed AMX-FP8 support into the LLVM compiler stack. This FP8 extension to the Advanced Matrix Extensions (AMX) is coming with next-gen [1]Diamond Rapids processors alongside other new ISA extensions .

AMX-FP8 was disclosed this month in Intel's updated public documentation as one of the new capabilities to be introduced with next-gen Xeon "Diamond Rapids" processors. AMX-FP8 allows for FP8 tiles with [2]Advanced Matrix Extensions .

AMX-FP8 with [3]Diamond Rapids complements the earlier Granite Rapids additions of the AMX-FP16 extension as well as AMX-COMPLEX that added support for half-precision floating point complex numbers. With Diamond Rapids FP8 input now rounds out the BF16 / INT8 / FP16 support for AMX. Since its introduction in Sapphire Rapids, AMX has shown [4]it can yield significant performance benefits for a variety of mostly AI-related software that is adapted to make use of the AMX instructions.

[5]This commit today to LLVM 20 Git adds the AMX-FP8 infrastructure and part of Intel's ongoing upstreaming of LLVM support for Diamond Rapids. The GCC upstreaming of new "-march=diamondrapids" ISA features also remains ongoing.



[1] https://www.phoronix.com/news/Intel-GCC-Diamond-Rapids-ISA

[2] https://www.phoronix.com/search/Advanced+Matrix+Extensions

[3] https://www.phoronix.com/search/Diamond+Rapids

[4] https://www.phoronix.com/review/intel-xeon-amx

[5] https://github.com/llvm/llvm-project/commit/8127162427c5f8c28d6292e1d4b4ce8a00b2d5a2#diff-8d728bd131fe53704180b3f8784f0688ba6f4ba6eafcf6469336573f26f10075R2