Graphcore and Hugging Face are teaming up to make state-of-the-art Transformer models run on Graphcore’s Intelligence Processing Units without developers needing to learn new code. The announcement came at the 2021 AI Hardware Summit, where Hugging Face unveiled its new Hardware Partner Program. Graphcore is a founding member.
The partnership centers on Optimum, a new open-source library that will host hardware-optimized models certified by Hugging Face. The first IPU-optimized models are expected later this year, covering applications from vision and speech to translation and text generation. Hugging Face CEO Clément Delangue framed the appeal bluntly: “Developers all want access to the latest and greatest hardware – like the Graphcore IPU, but there’s always that question of whether they’ll have to learn new code or processes. With Optimum and the Hugging Face Hardware Program, that’s just not an issue. It’s essentially plug-and-play.”
The numbers behind Hugging Face explain why Graphcore wants in. The Transformers library sees roughly 2 million downloads every month, and the platform has attracted more than 50,000 developers—the fastest adoption of an open-source project ever recorded, according to the company. For Graphcore, that’s a distribution channel directly into the hands of AI engineers who’ve largely defaulted to NVIDIA GPUs.
Graphcore’s pitch is architectural. Unlike GPUs, which use SIMD/SIMT designs, the IPU employs a massively parallel MIMD architecture with memory stacked directly on the silicon die. The company’s Poplar SDK already integrates with PyTorch, TensorFlow, Docker, and Kubernetes, so the Hugging Face partnership is less about basic compatibility and more about removing the last friction points for developers who want to experiment with non-GPU hardware. Graphcore has already demonstrated BERT-Large training benchmarks on IPUs that it says beat comparable GPU systems—though the company didn’t publish specific numbers in this announcement.
That silence on concrete performance figures is worth noting. “Dramatic benchmark results” is doing a lot of work without a single percentage point to back it up. Still, the strategic logic is sound: Hugging Face has become the default model repository for NLP practitioners, and Graphcore needs a way to reach those developers where they already work. If Optimum delivers genuinely plug-and-play IPU support, Graphcore might finally convert its architectural promises into actual market share.
