Accurate Models of AMD Matrix Cores
3 hours ago
- Matrix multipliers on recent GPUs do not conform to the IEEE 754 floating point standard.
- Features such as accumulator width, rounding, normalization, underflow/overflow, subnormals, and special inputs vary across vendors and architectures.
- Reproducibility of results across devices is impossible without documented implementation details.
- The study characterizes numerical behavior across three AMD GPU architectures: CDNA 1 (MI100), CDNA 2 (MI210/250), and CDNA 3 (MI300A/300X).
- Test vectors were designed to target numerical features for all supported input formats.
- MATLAB-based software models were developed for each architecture and validated for bit-level reproducibility using 10 million random test vectors.
- The models were used in two numerical applications to compare accuracy differences between AMD matrix cores and NVIDIA tensor cores.