AMD released the Typical Attained FLOPs (TAF) metric on October 6, 2026, saying it provides a standardized measure of sustained GEMM compute performance. It is the company's first performance update in the AI/ML category since the introduction of Peak FLOPs and Max-Achievable FLOPs (MAF).
AMD reported TAF results of 2,112 FP16, 3,072 BF16, and 7,168 unscaled FP8 sustained throughput on the MI325X GPU, measured across 10 representative GEMM shapes. That compares with earlier MAF results that focused on the highest-performing GEMM shape.
The TAF metric is built on a fixed set of 10 GEMM shapes derived from the DeepSeek V3 model and targets sustained compute performance in AI workloads. Availability begins with the MI325X GPU, initially for developers and data scientists.
"TAF provides a standardized measure of sustained GEMM compute performance rather than a direct measure of end-to-end application performance," said Ben Sander, AMD's AI/ML lead. The metric helps developers understand real-world performance across a range of GEMM sizes.
The announcement follows AMD's previous blogs on Peak FLOPs and MAF, which together with TAF provide a more complete view of GPU compute performance.
AMD did not say how TAF will be used in future applications, and the company raised the open question of whether TAF will be extended to other GPU models. The company plans to use the same 10-shape set for future GPU evaluations.
Source: amd
