Intel has removed AMX-TF32 from its ISA programming reference manual before the feature ever shipped in hardware. AMX-TF32 was a planned extension to Advanced Matrix Extensions (AMX) that would have natively supported NVIDIA's TensorFloat-32 format, offering FP32 range with FP16-like performance for AI/matrix workloads. The feature was previously confirmed for Diamond Rapids Xeon processors expected in 2027. Intel is now patching GCC to remove AMX-TF32 support, citing no actual hardware implementing the feature. This follows a similar fate to AMX-TRANSPOSE, which was also cancelled last year. Back-ported patches are needed to remove AMX-TF32 from GCC 15 and GCC 16.
121 Impressions