Accelerating ProtST on Intel Gaudi 2: 1.76x inference and 2.92x fine-tuning speedups

AI Impact Summary

ProtST has been re-architected to run on Intel Gaudi 2 using Optimum for Intel Gaudi, delivering near-linear scaling when using multiple accelerators. In benchmarks, Gaudi 2 achieves 1.76x faster inference than Nvidia A100 on ProtST-SubcellularLocalization and 2.92x faster fine-tuning for ProtST-ESM1b-for-sequential-classification on ProtST-BinaryLocalization, with distributed training scaling near-linearly across 4–8 Gaudi 2 devices. The work leverages models hosted on Hugging Face Hub and requires minimal code changes to port transformer-based scripts to Gaudi 2. This enables faster iteration cycles for protein design workflows and could reduce compute costs for large-scale inference and fine-tuning tasks.