AMD matrix cores do not follow the IEEE 754 floating point standard. Their features vary across architectures. This makes small matrix results unreproducible across devices. Researchers tested MI100, MI210/250, and MI300A/300X GPUs. They built MATLAB models for each architecture. Models were validated against hardware using 10 million random test cases. Models now match hardware outputs exactly. A proof-of-concept compared AMD and NVIDIA tensor core accuracy. The paper was submitted on 13 September 2026.

• 3 AMD GPU architectures
• 10 million random input vectors
• MI100, MI210/250, MI300A/300X GPUs