Hugging Face and Graphcore partner to optimize Transformers on IPU with Optimum integration

AI Impact Summary

Hugging Face and Graphcore are formalizing hardware-accelerated Transformers through the IPU-focused Hardware Partner Program. The Poplar SDK now integrates with PyTorch, TensorFlow, and standard deployment stacks (Docker, Kubernetes), enabling easy porting of models such as BERT and ViT to IPUs with minimal code changes. Optimum will host hardware-optimized, Graphcore-certified Transformer models, with the first IPU-optimized models appearing later this year. For engineering teams, this lowers the barrier to production-scale IPU deployment and could yield meaningful reductions in training and inference time.