Intel's Gaudi 2 fine-tunes ProtST protein model 2.92x faster than Nvidia A100

Intel and MILA have re-architected ProtST, the multi-modal protein language model that made waves at ICML 2023, and released it on the Hugging Face Hub. The move makes a model that has already racked up more than 40 citations in under a year far more accessible to researchers who don’t want to wrestle with custom infrastructure.

The performance numbers are the headline here. Running inference on the ProtST-SubcellularLocalization test set—2,772 amino acid sequences ranging from 79 to 1,999 residues—a single Gaudi 2 delivered 1.76x faster inference than an Nvidia A100 80GB PCIe at identical 0.44 accuracy. Fine-tuning tells an even starker story: on the binary localization task (membrane-bound vs. soluble), Gaudi 2 hit 2.92x the A100’s speed while matching the published accuracy of roughly 92.5%. The results held in full bfloat16 precision with batch size 1.

What makes this notable isn’t just the raw speedup. The Optimum for Intel Gaudi library means researchers port transformer-based scripts with minimal code changes—no CUDA-to-SYCL rewrites, no custom kernels. That’s the real barrier Intel is trying to break down. Hardware accelerators live and die by software ecosystem, and Gaudi 2’s pitch has always been that it slots into existing Hugging Face workflows. These ProtST benchmarks are concrete evidence that claim holds up.

Distributed scaling looks healthy too. The fine-tuning experiments show near-linear speedup when moving from one to four to eight Gaudi 2 accelerators, which matters for labs that need to iterate on larger protein datasets or longer sequences. The subcellular localization task is a standard benchmark in computational biology, but the zero-shot capability of ProtST-ESM-1b—outperforming few-shot classifiers without any task-specific training—suggests the model’s reach extends well beyond this single use case. Intel clearly sees biology as a growth market for its accelerator line, and these numbers give prospective buyers something concrete to weigh against Nvidia’s dominance. Whether price-per-token economics favor Gaudi 2 at scale remains an open question, but the performance gap is no longer a reason to dismiss it.