Graphcore

Graphcore. Graphcore is a British semiconductor company that designs and produces a specialized processor known as the Intelligence Processing Unit (IPU) for machine learning and artificial intelligence workloads. Founded in 2016, the company has attracted significant investment from prominent venture capital firms and strategic partners. Its technology is positioned as an alternative to GPUs from companies like Nvidia for accelerating AI research and data center applications.

History

Graphcore was founded in 2016 in Bristol by semiconductor industry veterans Nigel Toon and Simon Knowles, who had previously co-founded Icera, which was acquired by Nvidia. The company emerged from stealth mode in 2016 with the announcement of its novel processor architecture. In 2018, Graphcore secured a substantial Series D funding round led by Atomico and Sofina, which valued the company at over $1.7 billion, granting it unicorn status. The company established its headquarters in Bristol and later opened major offices in London, Cambridge, Palo Alto, Beijing, and Hsinchu. A significant partnership was formed with Microsoft in 2020 to offer its IPUs on the Azure cloud platform. However, by early 2024, facing intense market competition and financial challenges, Graphcore announced it was seeking new investors and considering a potential sale to a larger technology company.

Products

The company's primary product line is its Intelligence Processing Unit (IPU) systems. The first-generation Colossus MK1 IPU was launched in 2018, followed by the more powerful second-generation Colossus MK2 IPU in 2020. These processors are integrated into various hardware form factors, including the IPU-M2000 blade, which contains four MK2 IPUs, and the larger-scale IPU-POD systems that can scale to hundreds or thousands of interconnected IPUs for massive parallel computing tasks. Graphcore also provides a comprehensive software stack, Poplar, which is a graph-based framework designed to allow developers to program the IPU efficiently, supporting popular machine learning frameworks like PyTorch and TensorFlow. The company's products have been utilized for research in areas such as natural language processing, computer vision, and reinforcement learning.

Technology

The core technological innovation is the Intelligence Processing Unit, a manycore processor architecture specifically designed for the sparse, graph-like computational patterns inherent in machine learning workloads, unlike the SIMD architecture dominant in GPUs. Each IPU contains a large number of independent processor cores (tens of thousands) with substantial on-chip SRAM, known as In-Processor Memory, which aims to reduce latency and energy consumption by keeping model parameters on-chip. The Poplar software compiler explicitly extracts and maps the computational graph of an AI model directly onto the IPU's physical cores. This graph-native approach is intended to provide fine-grained parallelism and efficiency for both training and inference phases of complex models, including modern transformer architectures and foundation models.

Business and funding

Graphcore has been one of the most heavily funded tech startups in Europe. Early investors included Robert Bosch Venture Capital and Samsung Catalyst Fund. Major funding rounds were led by firms like Atomico, Sofina, BMW i Ventures, and Microsoft. The company also secured a strategic investment from AstraZeneca for life sciences applications. Despite raising over $700 million, Graphcore faced significant financial headwinds. In 2022, it reported substantial losses and a sharp decline in valuation during a funding round with Sequoia Capital. By 2024, the need for additional capital led to a public search for a buyer, with reported discussions with potential acquirers from various sectors, including other semiconductor design houses.

Competition and market position

Graphcore operates in the highly competitive market for AI accelerators, directly challenging the dominance of Nvidia and its CUDA ecosystem. Other key competitors include AMD with its Instinct series, Intel with Habana Labs and Gaudi, and Amazon Web Services with its Inferentia and Trainium chips. The company positioned its IPU technology for cutting-edge AI research and demanding high-performance computing applications, securing early evaluations with research institutions like the University of Oxford and the University of Cambridge. However, overcoming the entrenched software ecosystem of Nvidia proved a significant hurdle. Its market position was also affected by the broader semiconductor shortage and the capital-intensive nature of competing with established giants like TSMC for advanced manufacturing nodes.