Microscaling Data Formats for Deep Learning
This addresses the need for efficient data formats to lower costs in deep learning applications, representing a novel method for a known bottleneck rather than a foundational breakthrough.
The paper tackles the problem of reducing computational and storage costs in deep learning by evaluating Microscaling (MX) data formats, which combine per-block scaling with narrow floating-point and integer types, and demonstrates their practicality as a drop-in replacement for FP32 in AI inference and training, including training generative language models at sub-8-bit weights, activations, and gradients with minimal accuracy loss.
Narrow bit-width data formats are key to reducing the computational and storage costs of modern deep learning applications. This paper evaluates Microscaling (MX) data formats that combine a per-block scaling factor with narrow floating-point and integer types for individual elements. MX formats balance the competing needs of hardware efficiency, model accuracy, and user friction. Empirical results on over two dozen benchmarks demonstrate practicality of MX data formats as a drop-in replacement for baseline FP32 for AI inference and training with low user friction. We also show the first instance of training generative language models at sub-8-bit weights, activations, and gradients with minimal accuracy loss and no modifications to the training recipe.