Accurate Models of AMD Matrix Cores
The paper builds bit-level software replicas of AMD matrix-core behavior across CDNA 1, 2, and 3 GPUs.
The authors say AMD’s matrix multipliers do not follow IEEE 754 in documented, reproducible ways, and behavior varies by architecture. They designed targeted test vectors for supported input formats, then refined MATLAB models until they matched hardware on 10 million randomized input sets. The tested hardware includes MI100, MI210/250, and MI300A/300X GPUs. The models are also used to compare application-level numerical differences with NVIDIA tensor cores. HN · ArXiv's note
The authors say AMD’s matrix multipliers do not follow IEEE 754 in documented, reproducible ways, and behavior varies by architecture. They designed targeted test vectors for supported input formats, then refined MATLAB models until they matched hardware on 10 million randomized input sets. The tested hardware includes MI100, MI210/250, and MI300A/300X GPUs. The models are also used to compare application-level numerical differences with NVIDIA tensor cores. HN · ArXiv's note
score 5