AMD MI325X GPUs Reach 97 TCUPS in Protein Search Test

AMD’s Instinct MI325X GPUs processed up to 97 trillion cell updates per second in a protein search test, showing how the company’s accelerators can handle biological sequence analysis alongside other computing workloads.

The Sept. 21, 2026, benchmark used ROCm-enabled MMseqs2 to search the UniRef50 protein database with a single query sequence. It tested one, two, four and eight MI325X GPUs. AMD’s technical article describing the tools and results was published Oct. 14, 2026.

The MI325X outperformed Nvidia’s H100 in every tested configuration, while AMD’s MI300X performed comparably to or faster than the H100.

The two open-source tools cover different stages of sequence analysis. HIPSW++ runs exact Smith-Waterman alignment with affine gap penalties. MMseqs2-GPU filters out likely nonmatches before more computationally intensive alignment and can divide a database across multiple AMD GPUs in one server.

Evolutionary searches help provide biological data for models used in protein-fold prediction, gene annotation and protein design. Running search and later protein-folding workloads on the same accelerator can reduce data movement and simplify deployment.

The benchmark included data loading, initialization and preprocessing, but excluded database build time. Results may vary with system configurations.