[x86] AI Compute Extensions (ACE) Specification
5 hours ago
- #ML Workloads
- #x86 Extensions
- #Matrix Acceleration
- The document defines x86 extensions for accelerating computation, focusing on matrix multiplication kernels and reduced precision data formats relevant to machine learning workloads.
- ACE extensions introduce matrix multiplication primitives that augment AVX and scalar code, adding new capabilities such as ACE register state (tile and block scale registers), data processing operations using AVX register input, and data move operations between ACE and AVX registers.
- ACE enables tight integration between AVX vectors and ACE tile registers, combining high-density tile processing with AVX's comprehensive data processing features.
- In addition to matrix acceleration, dedicated format conversion operations are provided under the AVX10 framework.