MX-SAFE: New Microscaling Format for Efficient Deep Learning
There's a new format called MX-SAFE (MXSF) that's designed to improve cost efficiency in deep learning through quantization. It builds on the MX standard created by the Open Compute Project (OCP) back in 2022. MX-SAFE offers two adjustable modes: one with a wider mantissa (FP8 E2M5) and another for subnormal floating points (FP5 E3M2). This versatility supports both training and direct-cast inference, addressing the need for efficient computation in AI tasks.
Key facts
- MX-SAFE is a microscaling format for deep learning quantization.
- It builds on the MX format standardized by OCP in 2022.
- MX-SAFE uses two modes: FP8 E2M5 and FP5 E3M2.
- It supports both training and direct-cast inference.
- The format aims to reduce data size and cost.
- MX format shares an 8-bit exponent across multiple operands.
- MXINT focuses on high precision; MXFP on wider dynamic range.
- MX-SAFE is a versatile MXFP format.
Entities
Institutions
- Open Compute Project