# Graphcore

Graphcore Limited is a British semiconductor company that develops accelerators for AI and machine learning, notably its massively parallel Intelligence Processing Unit (IPU) which holds complete machine learning models on-chip. Founded in 2016, it was acquired by SoftBank Group in 2024.

Graphcore Limited is a British semiconductor company that develops accelerators for AI and machine learning. Its flagship product, the Intelligence Processing Unit (IPU), is a massively parallel processor designed to hold the entire machine learning model inside the chip, reducing the need for constant data movement between memory and compute. The company was founded in 2016 and became a notable player in the AI hardware space before being acquired by SoftBank Group in 2024.

The IPU architecture differs from traditional GPUs by using a large number of small processing cores, called tiles, each with its own local memory. This design allows for high-bandwidth, low-latency access to model parameters and activations, which is particularly suited to [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) workloads such as [large-language-model](https://www.wikiprompt.org/wiki/large-language-model) training and inference. Graphcore also developed the Poplar software stack to support popular machine learning frameworks like [TensorFlow](https://www.wikiprompt.org/wiki/tensorflow) and PyTorch.

## History

Graphcore was founded in 2016 by Simon Knowles and Nigel Toon, both veterans of the semiconductor industry. The company quickly attracted venture capital, with a first funding round led by Robert Bosch Venture Capital in autumn 2016, and additional backing from Samsung, Amadeus Capital Partners, C4 Ventures, Draper Esprit, Foundation Capital, and Pitango. In July 2017, a Series B round led by Atomico was followed by $50 million from Sequoia Capital. By December 2018, Graphcore closed a Series D of $200 million at a $1.7 billion valuation, making it a unicorn, with investors including 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. In early 2023, Meta Platforms acquired the AI networking technology team from Graphcore, though the company continued its core operations. In July 2024, SoftBank Group agreed to acquire Graphcore for around $500 million, a deal that was under review by the UK's Business Department's investment security unit.

## Products and Technology

Graphcore's first product announcement came in 2016 with the Poplar Software Stack, described as the world's first graph tool chain for machine intelligence. In July 2017, the company unveiled its first chip, the Colossus GC2, a 16 nm massively parallel, mixed-precision floating point processor that became available in 2018. The GC2 was packaged as two chips on a single PCI Express card, called the Graphcore C2 IPU, designed to perform the same role as a GPU in conjunction with standard machine learning frameworks. The device relies on scratchpad memory for performance rather than traditional cache hierarchies.

In July 2020, Graphcore introduced its second-generation processor, the GC200, built using [TSMC](https://www.wikiprompt.org/wiki/tsmc)'s 7nm FinFET 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 of a GC200 die bonded face-to-face to a power-delivery die, enabling higher clock rates at lower core voltage. The company also outlined a long-term goal called the 'Good machine', named after I.J. Good, which would support AI models with more parameters than the human brain has synapses.

## Architecture Details

Both the GC2 and GC200 chips support 6 threads per tile, yielding 7,296 and 8,832 threads respectively. The architecture uses MIMD (Multiple Instruction, Multiple Data) parallelism, with distributed, local memory as the 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, and single-precision FP32 at lower performance. Code and data executed locally must fit within a tile, but message-passing allows use of all on-chip or off-chip memory, and software for AI makes this transparent, including PyTorch support.

## Market Position and Impact

Graphcore positioned itself as a challenger to [Nvidia](https://www.wikiprompt.org/wiki/nvidia) in the AI accelerator market, focusing on a unique architecture that differs from both GPUs and other AI chips like those from [Cerebras](https://www.wikiprompt.org/wiki/cerebras) or [Groq](https://www.wikiprompt.org/wiki/groq). Its IPU was particularly aimed at training and inference for [neural-network](https://www.wikiprompt.org/wiki/neural-network) models, with early adopters including Microsoft Azure and various research institutions. The acquisition by SoftBank in 2024 was seen as a strategic move to bolster SoftBank's AI ambitions, though the deal faced regulatory scrutiny in the UK. Graphcore's technology continues to be used in cloud and enterprise settings, and its software stack remains an active area of development.

## See Also

- AI accelerator
- Massively parallel processor array

---
Source: https://www.wikiprompt.org/wiki/graphcore-ipu
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-05T14:06:55.765865+00:00