Graphcore: Building the Processor for Machine Intelligence

In a converted warehouse district near Bristol's floating harbour, a company has spent the better part of a decade attempting something few organisations dare: designing a new kind of processor from scratch. Graphcore, founded in 2016 by Nigel Toon and Simon Knowles, set out to build silicon purpose-made for artificial intelligence workloads, rather than repurposing chips originally designed for graphics rendering or general computation. The result is the Intelligence Processing Unit, or IPU, a chip architecture that takes a fundamentally different approach to the way AI models are trained and deployed.

A Different Kind of Chip

The semiconductor industry is not short of incumbents. Nvidia's dominance in AI accelerators is well documented, and a handful of well-funded startups have tried and failed to challenge it. What sets Graphcore apart is the depth of its architectural conviction. The IPU was not designed as an incremental improvement on existing GPU designs. Instead, it was built around the principle that machine intelligence workloads, particularly those involving sparse and irregular computation, demand a processor with massive parallelism, distributed on-chip memory, and a communication fabric that allows thousands of independent processing threads to exchange data efficiently.

This is not a minor engineering distinction. It shapes everything about how the company operates, from the compiler stack to the system-level software, from the developer tools to the research partnerships. At Graphcore, hardware and software are deeply intertwined. One cannot function without the other, and the teams that build them work in close collaboration.

Bristol Roots, Global Reach

Bristol has long been an underappreciated centre for chip design in the UK. The city's semiconductor ecosystem traces back decades, with deep roots in companies like Inmos and STMicroelectronics. Graphcore drew on this local talent pool from the start, and the city remains its engineering heartland. The company also has offices in London, Oslo, Gdańsk, and locations across Asia and North America, but Bristol is where the core architecture and much of the systems software work takes place.

The choice of Bristol was not incidental. The city offered a rare concentration of hardware engineers, compiler specialists, and systems programmers, people who understand what it means to design at the transistor level and then build the software layers that make a chip useful. This talent density has been central to Graphcore's ability to iterate quickly on a product that requires expertise spanning analogue design, digital logic, operating systems, graph compilers, and machine learning frameworks.

The Poplar Software Stack

A processor is only as useful as the software that programs it. Graphcore recognised this early and invested heavily in Poplar, its proprietary software stack. Poplar sits between the hardware and the frameworks that researchers and engineers use daily, such as PyTorch and TensorFlow. It includes a graph compiler that maps computation onto the IPU's architecture, a set of optimised libraries, and tools for profiling and debugging performance at a granular level.

Building a compiler for a novel architecture is one of the harder problems in computer science. It requires not only deep knowledge of code generation and optimisation but also an intimate understanding of the hardware's memory hierarchy, communication topology, and execution model. For engineers who enjoy working at this level of abstraction, Graphcore offers a rare opportunity: the chance to shape how an entirely new class of processor is programmed.

"We are not adapting someone else's architecture. We are defining what machine intelligence hardware should look like, and then writing the software to prove it."

Research and Partnerships

Graphcore has collaborated with a broad range of institutions and companies. Its IPU systems have been used in academic research spanning genomics, molecular dynamics, natural language processing, and computer vision. The company has published work demonstrating competitive performance on large-scale AI benchmarks and has explored areas where the IPU's architecture offers distinct advantages, such as models with irregular sparsity or those requiring fine-grained communication between processing elements.

These partnerships are not purely commercial exercises. They feed back into the product roadmap, helping the engineering teams understand where the hardware and software need to evolve. For employees, this means exposure to real-world AI workloads and the chance to see how architectural decisions translate into measurable outcomes.

Culture and Working Life

Graphcore's culture reflects its engineering-led origins. Decisions tend to be driven by technical merit rather than hierarchy, and there is an expectation that people at all levels will engage deeply with the problems they are solving. The company is large enough to have meaningful structure, with dedicated teams for hardware, systems software, applications, and developer relations, but small enough that individual contributions remain visible.

Bristol itself adds to the appeal. The city has a strong creative and technological identity, a manageable cost of living compared to London, and a quality of life that consistently ranks among the highest in the UK. For engineers relocating from elsewhere, the combination of meaningful technical work and a liveable city is a compelling proposition.

Looking Ahead

The AI hardware landscape continues to shift rapidly. Demand for compute is growing faster than almost anyone predicted, and the question of what architectures will best serve the next generation of models remains genuinely open. Graphcore occupies a distinctive position in this conversation: a company that has built a novel processor, a full software stack, and a body of research demonstrating what that processor can do. Whether the company ultimately reshapes the market or serves as a catalyst for broader architectural innovation, the work being done in Bristol is serious, technically ambitious, and consequential.