Block math

• Block computation similar to OCP MX
• Architecture-agnostic by leveraging INT8 / FP8 dot intrinsics
• Up to ~30% throughput improvement over BF16

Installation

git clone https://github.com/lightb0x/block-math.git
cd block-math
python3 -m pip install -e .

Supports

• Datatypes: float8_e5m2, int8
• Scale : E8 inside float32
• Architectures
  • Ampere and newer for int8
  • Ada/Hopper and newer for fp8
  • Blackwell not tested