Matrix multipliers on recent GPUs do not conform to IEEE 754, with differences across vendors and architectures in accumulator width, rounding, normalization, subnormals, and special inputs. This prevents reproducibility across devices via software.

Implementation details are undocumented, complicating interpretation of discrepancies. Researchers characterize numerical behavior across three AMD GPU architectures: CDNA 1, CDNA 2, and CDNA 3, using MI100, MI210/250, and MI300A/300X GPUs.

Test vectors target numerical features for all supported input formats, with derivations provided. MATLAB-based software models are developed for each architecture and validated for bit-level reproducibility against hardware using 10 million randomized test sets. A previously developed iterative refinement technique ensures model accuracy.

As proof of concept, the models are used in two numerical applications, quantifying application-level accuracy differences between AMD matrix cores and NVIDIA tensor cores.