Artificial intelligence (AI) semiconductor company Graphcore has joined hands with Hugging Face, a leader in the open-source AI field, to accelerate machine intelligence innovation.

Graphcore announced on the 17th that it has entered into a partnership with Hugging Face and plans to participate in Hugging Face's Hardware Partner Program to realize the two companies' shared goal of accelerating machine intelligence innovation.

This enables developers to implement SOTA (State-of-the-Art) Transformer models optimized for Graphcore IPUs (Intelligence Processing Units) at production scale with minimal coding through the HuggingFace hardware partner program.

The IPU is the processor that forms the basis of Graphcore's data center computing system, the IPU-POD, and is designed to support the unique computing specifications required specifically for AI and machine learning (ML). It features characteristics such as granular parallel processing, low-precision computation, and sparse processing capabilities.

Instead of a SIMD/SIMT architecture like GPUs, the Graphcore IPU utilizes a massively parallel MIMD architecture with ultra-high bandwidth memory adjacent to the processor cores directly on the silicon die. This design provides groundbreaking performance and efficiency for running today's most widely used models, such as BERT and EfficientNet, as well as for exploring next-generation AI applications.

Graphcore has emphasized the importance of software in maximizing IPU capabilities. Since its inception, Graphcore has co-designed all processors and its proprietary Poplar SDK software, ensuring compatibility with the third-party systems most widely used by developers. Developers can easily port models to other computing platforms to leverage the performance benefits of the IPU.

HuggingFace already possesses hundreds of Transformers, including CamemBERT for the French language and ViT, which applies NLP insights to computer vision. Through this partnership, HuggingFace is expected to accelerate AI innovation by combining its best-in-class Transformer toolset with Graphcore's cutting-edge AI hardware.

Developers will be able to utilize hardware-optimized models certified by HuggingFace using Optimus, a new open-source library and toolkit. The two companies are jointly developing these models through close collaboration and plan to unveil the first model optimized for Graphcore IPUs in late 2021. Furthermore, the models under development are expected to be applied to a wide range of applications, including vision, speech, translation, and text generation.