Graphcore is a British semiconductor company headquartered in Bristol, England, that develops accelerators for artificial intelligence and machine learning. Founded in 2016 by Simon Knowles and Nigel Toon, the company is best known for its Intelligence Processing Unit (IPU), a massively parallel processor designed to hold entire machine learning models in on-chip memory. In July 2024, Graphcore was acquired by SoftBank Group for approximately $500 million, after facing commercial challenges in a market dominated by Nvidia and other established players.
History
Graphcore was founded in 2016 by Simon Knowles and Nigel Toon, both veterans of the semiconductor industry. Knowles had previously co-founded ARM Holdings spin-off Element 14 and chip designer Icera, while Toon had led XMOS and held senior roles at Altera. The company aimed to create a new class of processor tailored specifically for AI workloads, diverging from the GPU-centric approach that dominated the field.
In autumn 2016, Graphcore secured its first funding round led by Robert Bosch Venture Capital, with participation from Samsung Electronics, Amadeus Capital Partners, C4 Ventures, Draper Esprit, Foundation Capital, and Pitango. This early backing allowed the company to expand its engineering team and accelerate development of its IPU architecture.
In July 2017, Graphcore raised a Series B round led by Atomico, followed a few months later by an additional $50 million investment from Sequoia Capital. This capital injection valued the company at over $300 million and funded the tape-out of its first silicon.
In December 2018, Graphcore closed a Series D round of $200 million, bringing its valuation to $1.7 billion and earning it unicorn status. The round included investments from Microsoft, Samsung, and Dell Technologies. This funding supported the commercial launch of its first products and the expansion of its software ecosystem.
On 13 November 2019, Graphcore announced that its Graphcore C2 IPU cards were available for preview on Microsoft Azure, marking a significant milestone in cloud deployment. The partnership with Microsoft provided early customers with access to IPU-powered instances for AI research and development.
In early 2023, Meta Platforms acquired Graphcore's AI networking technology team, a move that saw several engineers join Meta's infrastructure group. This acquisition was part of Meta's broader effort to build custom silicon for its data centers, though Graphcore retained its core IPU business and remaining staff.
In July 2024, SoftBank Group agreed to acquire Graphcore for around $500 million. The deal, which was subject to regulatory review by the UK government's investment security unit, marked the end of Graphcore's independence. SoftBank's acquisition was seen as a strategic move to bolster its AI chip portfolio, which already included ARM Holdings and Nvidia shares.
Products
Graphcore's primary product line is the Intelligence Processing Unit (IPU), a processor designed from the ground up for AI and machine learning workloads. Unlike traditional GPUs, which are optimized for graphics and later adapted for AI, the IPU architecture prioritizes parallel processing and on-chip memory bandwidth.
In 2016, Graphcore announced the Poplar Software Stack, the first graph-based toolchain designed specifically for its IPU. Poplar allows developers to map machine learning models onto the IPU's many cores, abstracting away the complexities of parallel programming.
In July 2017, Graphcore unveiled its first chip, the Colossus GC2, a 16 nm processor with 1,216 cores and 256 KB of in-processor memory per core. The GC2 was packaged on a PCI Express card called the Graphcore C2, which contained two GC2 chips, providing a total of 2,432 cores and 512 MB of on-chip memory. The C2 card was designed to be inserted into standard servers, similar to GPUs, and supported popular machine learning frameworks such as TensorFlow and PyTorch.
In July 2020, Graphcore introduced its second-generation processor, the Colossus GC200, manufactured using TSMC's 7 nm process. The GC200 featured 1,472 cores, 900 MB of in-processor memory, and 59 billion transistors, making it one of the largest chips ever produced at the time. The GC200 was offered in the Mk2 IPU-POD systems, which could be scaled from a single card to large clusters with thousands of IPUs.
In 2022, Graphcore and TSMC collaborated on the Bow IPU, a 3D-packaged version of the GC200 that bonded a power-delivery die to the processor die. This advanced packaging technique allowed for higher clock speeds while reducing power consumption, improving overall performance per watt.
Both the GC2 and GC200 support 6 threads per core, enabling the GC2 to run 7,296 threads and the GC200 to run 8,832 threads simultaneously. The IPU's architecture is based on MIMD (Multiple Instruction, Multiple Data) parallelism, with each core executing independent instructions. Unlike GPUs, which rely on shared memory hierarchies, the IPU uses distributed local memory, with each core having its own dedicated memory block. This design eliminates the need for a separate cache hierarchy, as all data resides in on-chip memory, reducing latency and power consumption.
The IPU supports IEEE 754 half-precision (FP16) floating-point arithmetic with stochastic rounding, which is particularly suited for AI training. It also supports single-precision (FP32) at lower performance for applications requiring higher numerical accuracy. The processor is designed to execute entire machine learning models within its on-chip memory, avoiding the need for frequent off-chip data transfers that bottleneck traditional architectures.
Graphcore's software stack, including Poplar and the PopART library, provides tools for model compilation, optimization, and execution. The company also developed PopDist, a distributed training framework, and PopSparse, a library for sparse matrix operations. These tools aim to make it easier for developers to port existing models to the IPU and to exploit its unique capabilities.
See also
References
- "Graphcore - About Us". Graphcore. Retrieved 2025-01-01.
- "Graphcore raises $200M at $1.7B valuation". TechCrunch. 2018-12-18.
- "Meta acquires Graphcore's AI networking team". The Register. 2023-03-10.
- "SoftBank to acquire Graphcore". Financial Times. 2024-07-11.
- "Graphcore unveils Colossus GC200". AnandTech. 2020-07-15.
- "Graphcore and TSMC announce Bow IPU". Graphcore Press Release. 2022-05-25.