ProtST protein language model accelerates on Intel Gaudi 2 — 1.76x inference, 2.92x fine-tuning

AI Impact Summary

Intel Gaudi 2 accelerators deliver 1.76x faster ProtST inference and 2.92x faster fine-tuning for ProtST-ESM-1b versus a baseline (A100) setup, showing clear throughput gains for protein design workflows. The effort uses bfloat16 precision, single-batch inference, and near-linear scaling when distributing across 4–8 Gaudi 2 cards, with the Optimum for Intel Gaudi library enabling porting of Hugging Face transformer scripts. Researchers can reproduce results using the ProtST-SubcellularLocalization and ProtST-BinaryLocalization datasets and the provided scripts, signaling practical paths to faster experimentation on protein localization tasks.