Intel Removing AMX-TRANSPOSE From The GCC Compiler
One year ago updated Intel documentation noted AMX-TRANSPOSE as one of the new ISA additions for Diamond Rapids. But in updated Intel architecture documentation last month, it oddly removed all references to AMX-TRANSPOSE. Confirming that the Advanced Matrix Extensions (AMX) addition for TRANSPOSE is now dead, an Intel engineer posted a patch to remove AMX-TRANSPOSE from the GCC compiler.
Intel documentation in September removed references to AMX-TRANSPOSE without further elaborating. This was to be an addition with Xeon Diamond Rapids CPUs for transposing a matrix using AMX. But after already plumbing the support into the GCC compiler, the feature is now being removed ahead of next-generation Diamond Rapids processors launching.
This patch posted removes the AMX-TRANSPOSE code from GCC, signifying that Intel is looking to kill off the feature entirely as opposed to iterating on it or just pushing it back from Diamond Rapids:
AMX-TRANSPOSE should be removed in GCC 16 while mailing list communication is suggesting that for GCC 15 the easier route would be just dropping AMX-TRANSPOSE from being part of the Diamond Rapids target.
This isn't the first time in recent memory that Intel wires up a new compiler feature only to walk it back. Earlier this year they did the same around AVX10 256-bit support being dropped thankfully in making AVX10 512-bit mandatory for future Intel CPUs.
Intel documentation in September removed references to AMX-TRANSPOSE without further elaborating. This was to be an addition with Xeon Diamond Rapids CPUs for transposing a matrix using AMX. But after already plumbing the support into the GCC compiler, the feature is now being removed ahead of next-generation Diamond Rapids processors launching.
This patch posted removes the AMX-TRANSPOSE code from GCC, signifying that Intel is looking to kill off the feature entirely as opposed to iterating on it or just pushing it back from Diamond Rapids:
"AMX-TRANSPOSE is removed from ISE059. Since there is no actual hardware, we choose to directly remove it in GCC 16 and backport to GCC 15 to ease maintainence effort."
AMX-TRANSPOSE should be removed in GCC 16 while mailing list communication is suggesting that for GCC 15 the easier route would be just dropping AMX-TRANSPOSE from being part of the Diamond Rapids target.
This isn't the first time in recent memory that Intel wires up a new compiler feature only to walk it back. Earlier this year they did the same around AVX10 256-bit support being dropped thankfully in making AVX10 512-bit mandatory for future Intel CPUs.