Bit-Accurate Modeling of GPU Matrix Multiply-Accumulate Units: Demystifying Numerical Discrepancy and Accuracy
The work proposes a closed-loop feature probing framework to build bit-accurate arithmetic models for GPU matrix multiply-accumulate units (MMAUs). This allows researchers to diagnose numerical errors across different hardware platforms by systematically testing and refining models, revealing specific precision bottlenecks in instruction-level MMA operations.