
AMD and Intel release the ACE specification, a shared AI instruction set for x86
AMD and Intel's x86 Ecosystem Advisory Group released AI Compute Extensions (ACE) v1, a shared matrix-multiplication specification for both vendors' future CPUs covering 11 data formats — FP8, BF16, INT8, FP16, FP32, and several Microscaling variants — with PyTorch, TensorFlow, and NumPy integrations planned. The spec is downloadable now, but ACE-compatible hardware isn't expected before 2028: AMD hasn't mentioned it for Zen 6, and Intel hasn't for Nova Lake.
Source: heise.de ↗
ACE primarily deals with matrix multiplications with AI data formats for inferencing with quantized weights.
heise.de article
Why this matters
- → x86 gains unified AI acceleration standard, reducing fragmentation.
- → ARM's SME2 now faces direct x86 competition in matrix ops.
- → 2028+ timeline gives GPU vendors two-year buffer before hardware lands.
x86 fights back on AI