AMD Instinct MI100 product page GPU Specifications table

AMD's own labelling convention, established here and used as the reference for every AMD record: a row reading 'Peak X Performance' is DENSE, and a separate row reading 'Peak X Performance with Structured Sparsity' carries exactly 2x the dense value. On MI100 AMD prints 'Peak Single Precision (FP32) 23.1 TFLOPs' AND 'Peak Single Precision Matrix (FP32) 23.1 TFLOPs' - the same number twice - while ROCm's gpu-arch reference gives Vector FP32 128 FLOPS/clock/CU (23.1) and Matrix FP32 256 (46.1). AMD's matrix row is therefore internally inconsistent with its own architecture reference. AMD also prints 'Peak INT4 92.3 TOPs' and 'Peak INT8 92.3 TOPs' and 'Peak bfloat16 92.3 TFLOPs' - all equal, and all half the 184.6 FP16 Matrix figure.