Accurate Models of AMD Matrix Cores
Researchers have developed accurate software models of AMD GPU matrix cores (used in MI100, MI210/250, and MI300 series) that characterize their numerical behavior across different architectures and achieve bit-level reproducibility with hardware. Since AMD matrix multipliers deviate from IEEE 754 floating-point standards in ways that vary by architecture—affecting accumulator width, rounding behavior, and handling of special values—these models enable reproducibility studies and numerical comparisons that were previously impossible. The models were validated against 10 million randomized test cases and demonstrate application-level differences between AMD and NVIDIA tensor cores.