Graphcore

What Graphcore is

Graphcore is a British semiconductor company that designs processors for artificial intelligence. It was founded in 2016 and is headquartered in Bristol, England^[2]. The company describes itself as the inventor of the Intelligence Processing Unit (IPU), a processor designed for current and next-generation AI workloads^[1]. Where much of the industry takes general-purpose processors or graphics chips and adapts them to machine learning, Graphcore's claim is that its hardware was designed around the shape of AI computation from the start.

Origins

Graphcore was co-founded by Nigel Toon, its chief executive, and Simon Knowles, its chief technology officer^[2]. The company has kept its base in Bristol, where it remains headquartered^[2].

How the IPU differs

The IPU architecture relies on massively parallel processing and scratchpad memory rather than a conventional cache hierarchy^[2]. The difference is worth spelling out. In an ordinary processor, hardware cache logic decides which pieces of data to keep close to the compute units, and software can only influence that decision indirectly. Scratchpad memory works the other way round: it is a pool of fast on-chip memory that software addresses and manages explicitly, so the compiler, not the hardware, decides what sits where. Pair that with many cores working in parallel on the same problem, and you get a design aimed at the workload AI models produce, in which the same operations are applied to large blocks of data at once.

Second-generation hardware and systems

Graphcore's second-generation IPU, the GC200, was announced in July 2020 and built on TSMC's 7-nanometer process^[2]. It carries 59 billion transistors, 1,472 processor cores and 900 MB of on-chip SRAM^[2]. That last number is unusual for a processor: a large pool of memory on the chip itself keeps model data near the cores instead of sending it back and forth to external memory. At the system level, the company says its IPU-POD data centre systems can run large models across multiple IPUs, or share compute between different users and workloads^[1].

Funding and finances

Graphcore drew substantial backing from outside investors. In December 2018 it raised US$200 million in a Series D round at a US$1.7 billion valuation, with Microsoft, Samsung and Dell Technologies among the investors^[2]. The commercial numbers were less comfortable. For 2022 the company reported US$2.7 million of revenue and a net loss of about US$205 million^[2]. The distance between those two figures shows what it costs to develop a new processor architecture and the software stack that has to go with it, and how long it takes for that work to turn into sales.

Joining SoftBank

In July 2024 Graphcore became a wholly owned subsidiary of SoftBank Group, keeping its name and its Bristol headquarters^[1]. The company described the move as a step toward building the next generation of AI compute^[1].

Where it sits

Graphcore belongs to the group of companies that tried to build a chip business specifically around AI rather than around general-purpose computing. Its public case rests on two ideas: an architecture tuned for parallel, data-heavy workloads, and systems that let many chips be pooled to run large models or serve several users at once^[1][2]. Whether that approach pays off commercially is a separate question from whether it is technically distinctive, and the company's own reported results give the reader both sides to weigh^[2].