Graphcore Limited is a British semiconductor company that designs accelerators for artificial intelligence and machine learning workloads. The company is best known for its Intelligence Processing Unit (IPU), a massively parallel processor that holds the entire machine learning model within the chip itself, unlike conventional graphics processing units (GPUs) that rely on external memory. Graphcore's technology targets both training and inference of neural networks, aiming to provide an alternative to GPU-based systems for AI computing.
Founded in 2016, Graphcore grew rapidly through multiple funding rounds, reaching a valuation of $1.7 billion by December 2018. The company's products, including the Colossus GC2 and GC200 processors, have been used in cloud and enterprise settings. In July 2024, SoftBank Group agreed to acquire Graphcore for approximately $500 million, a deal that was under regulatory review in the United Kingdom.
History
Graphcore was founded in 2016 by Simon Knowles and Nigel Toon, both veterans of the semiconductor industry. Knowles had previously co-founded Icera, a baseband processor company acquired by NVIDIA, while Toon had led several chip startups. The company set out to design processors specifically for machine intelligence, a niche that was then dominated by GPUs.
In the autumn of 2016, Graphcore secured its first major funding round, led by Robert Bosch Venture Capital. Other investors included Samsung, Amadeus Capital Partners, C4 Ventures, Draper Esprit, Foundation Capital, and Pitango. This early backing allowed the company to begin developing its IPU architecture and software stack.
A Series B round followed in July 2017, led by Atomico, with additional funding from Sequoia Capital a few months later. By December 2018, Graphcore closed a Series D round of $200 million at a valuation of $1.7 billion, making it a "unicorn" (a private company valued at over $1 billion). Investors in this round included Microsoft, Samsung, and Dell Technologies.
In November 2019, Graphcore announced that its C2 IPU cards were available for preview on Microsoft Azure, marking a significant cloud deployment. The company continued to expand its product line and partnerships, but also faced challenges in a competitive market dominated by NVIDIA.
In early 2023, Meta Platforms acquired the AI networking technology team from Graphcore, a move that was seen as a strategic divestment. The financial terms were not disclosed, but it allowed Graphcore to focus on its core processor business.
In July 2024, SoftBank Group agreed to acquire Graphcore for around $500 million. The deal was under review by the UK's Business Department's investment security unit, reflecting national security concerns in the semiconductor sector. The acquisition was expected to provide Graphcore with additional resources to continue its development efforts.
Products and Technology
Graphcore's primary product line is based on the Intelligence Processing Unit (IPU), a processor designed from the ground up for AI workloads. The IPU architecture differs significantly from GPUs: it uses a massively parallel array of cores, each with its own local memory, and the entire machine learning model is stored on-chip. This design reduces the need for data movement between memory and processor, which is a major bottleneck in traditional architectures.
The first IPU chip, called Colossus GC2, was announced in July 2017. It was a 16 nm massively parallel, mixed-precision floating point processor that became available in 2018. The GC2 was packaged with two chips on a single PCI Express card, known as the Graphcore C2 IPU. This card was designed to perform the same role as a GPU in conjunction with standard machine learning frameworks such as TensorFlow. The device relied on scratchpad memory for its performance rather than traditional cache hierarchies.
In July 2020, Graphcore introduced its second-generation processor, the GC200, built using TSMC's 7nm FinFET manufacturing process. The GC200 is a 59 billion transistor, 823 square-millimeter integrated circuit with 1,472 computational cores and 900 Mbyte of local memory. In 2022, Graphcore and TSMC presented the Bow IPU, a 3D package that bonds a GC200 die face-to-face with a power-delivery die, allowing for higher clock rates at lower core voltages.
Both the GC2 and GC200 chips use 6 threads per tile, for a total of 7,296 and 8,832 threads respectively. They employ MIMD (Multiple Instruction, Multiple Data) parallelism and have distributed, local memory as their only form of memory on the device, except for registers. The older GC2 chip has 256 KiB per tile, while the newer GC200 has about 630 KiB per tile, arranged into islands of 4 tiles, which are then arranged into columns. Latency is best within a tile.
The IPU supports IEEE FP16 with stochastic rounding, as well as single-precision FP32 at lower performance. Code and data executed locally must fit in a tile, but with message-passing, all on-chip or off-chip memory can be used. Software for AI makes this transparently possible, for example, through PyTorch support.
Software Stack
In 2016, Graphcore announced the world's first graph tool chain designed for machine intelligence, called the Poplar Software Stack. Poplar is a complete software development kit that includes a graph compiler, a runtime, and libraries for building and deploying AI models. It abstracts away the complexities of the IPU hardware, allowing developers to write high-level code in frameworks like TensorFlow and PyTorch.
Poplar's graph-based approach is well-suited to the IPU's architecture, which is itself a graph of processing tiles. The compiler maps the computational graph of a neural network onto the IPU's tiles, optimizing for memory usage and communication. The runtime manages the execution of the graph on the hardware, including scheduling and data movement.
The software stack also includes PopART, a library for inference, and PopNN, a library for building neural networks. These tools are designed to be interoperable with existing AI frameworks, making it easier for developers to adopt IPU technology.
Market Position and Competition
Graphcore operates in the highly competitive field of AI accelerators, where NVIDIA has historically dominated with its GPUs. Other competitors include AMD with its Instinct line, Intel with its Gaudi processors, and startups like Groq and SambaNova. Graphcore's IPU differentiates itself by focusing on in-memory computing and a massively parallel architecture, which can offer advantages for certain types of models, particularly those with large memory footprints.
The company has targeted both cloud and edge markets. Its processors have been deployed in data centers through partnerships with cloud providers like Microsoft Azure and Oracle Cloud. Graphcore has also worked with research institutions and enterprises to accelerate AI workloads in areas such as natural language processing, computer vision, and scientific computing.
Despite its technological innovations, Graphcore has faced challenges in scaling its business. The AI accelerator market is capital-intensive, and the company has struggled to compete with NVIDIA's ecosystem and software maturity. The acquisition by SoftBank is seen as a way to provide financial stability and access to broader resources.
Future Directions
Graphcore has stated its ambition to build a "Good machine," named after I.J. Good, a British mathematician who worked with Alan Turing. The goal is to enable AI models with more parameters than the human brain has synapses, which would require significant advances in both hardware and software. The company continues to develop its IPU architecture, with plans for future generations that could offer even higher performance and efficiency.
The acquisition by SoftBank is expected to accelerate these efforts. SoftBank has a history of investing in AI and semiconductor companies, including Arm Holdings, and could provide Graphcore with the capital needed to expand its manufacturing and global reach. The deal's regulatory approval is still pending, but if completed, it could mark a new chapter for the British chip designer.