Hugging Face and Graphcore partner for IPU-optimized Transformers

Speaking at the 2021 AI Hardware Summit, Hugging Face announced the launch of their new Hardware Partner Program, including device-optimized models and software integrations. Here, Graphcore - creators of the Intelligence Processing Unit (IPU) and a founding member of the program – explain how their partnership with Hugging Face will allow developers to easily accelerate their use of state-of-the-art Transformer models.

Graphcore and Hugging Face are two companies with a common goal – to make it easier for innovators to harness the power of machine intelligence.

Hugging Face’s Hardware Partner Program will allow developers using Graphcore systems to deploy state-of-the-art Transformer models, optimised for our Intelligence Processing Unit (IPU), at production scale, with minimum coding complexity.

What is an Intelligence Processing Unit?

IPUs are the processors that power Graphcore’s IPU-POD datacenter compute systems. This new type of processor is designed to support the very specific computational requirements of AI and machine learning. Characteristics such as fine-grained parallelism, low precision arithmetic, and the ability to handle sparsity have been built into our silicon.

Instead of adopting a SIMD/SIMT architecture like GPUs, Graphcore’s IPU uses a massively parallel, MIMD architecture, with ultra-high bandwidth memory placed adjacent to the processor cores, right on the silicon die.

This design delivers high performance and new levels of efficiency, whether running today’s most popular models, such as BERT and EfficientNet, or exploring next-generation AI applications.

Software plays a vital role in unlocking the IPU’s capabilities. Our Poplar SDK has been co-designed with the processor since Graphcore’s inception. Today it fully integrates with standard machine learning frameworks, including PyTorch and TensorFlow, as well as orchestration and deployment tools such as Docker and Kubernetes.