Accelerating Protein Language Model ProtST on Intel Gaudi 2

Source: HuggingFace Date: 2024-07-03 URL: https://huggingface.co/blog/intel-protein-language-model-protst

Summary

Integration tutorial: Running and fine-tuning ProtST (MILA/Intel protein language model, ESM-1b base) on Intel Gaudi 2 via Optimum for Intel Gaudi. Task: subcellular protein localization prediction. Inference: Gaudi 2 is 1.76x faster than A100 80GB at identical accuracy (0.44). Fine-tuning on binary localization: 2.92x faster than A100, matching published ~92.5% accuracy. Near-linear scaling observed at 4-8 Gaudi 2 accelerators.

Implications

Open-weights ecosystem health. Gaudi 2 at 2.92x fine-tuning speedup over A100 is a meaningful hardware result for bioinformatics teams — protein language models are computationally intensive and labs doing structure prediction or protein design at scale have concrete cost reasons to evaluate Intel accelerators. The Optimum for Intel Gaudi abstraction layer requiring minimal code changes lowers the migration barrier.

HF as open-source ML hub. Intel publishing scientific domain-specific benchmarks (protein models on Gaudi) through the HF blog continues to establish HF as the distribution channel for hardware vendors’ ML credibility claims. The pattern: library + benchmark + tutorial on HF = legitimate performance claim in the ML community’s working vocabulary.