Graphcore

Graphcore is one of the most technically distinctive and intellectually original companies in the AI hardware space — the Bristol-based semiconductor company that took a fundamental look at the computational requirements of machine learning and decided that existing processor architectures, all of which adapted hardware originally designed for other purposes to AI workloads rather than designing from the ground up for them, were leaving an enormous performance opportunity on the table. Founded in 2016 by Simon Knowles and Nigel Toon, Graphcore created the Intelligence Processing Unit — a new category of processor designed from first silicon to match the specific computational patterns, memory access characteristics, and parallelism requirements of modern machine intelligence algorithms. The Graphcore IPU's architecture is fundamentally different from a GPU in ways that matter profoundly for AI workloads. Where a GPU organizes computation around a small number of very large, very powerful processing cores surrounded by a hierarchical memory system optimized for graphics rendering, the IPU organizes computation around thousands of independent processors, each with its own local SRAM memory, connected by a high-bandwidth low-latency fabric. This bulk synchronous parallel model maps exceptionally well to the graph-structured computations at the heart of neural network training and inference, enabling the IPU to process the fine-grained, irregular data access patterns of modern AI models with efficiency that GPU architectures struggle to match. Graphcore's Bow IPU systems and Bow Pod infrastructure deliver these architectural advantages in production hardware that research institutions, pharmaceutical companies, financial services firms, and AI laboratories have deployed for demanding machine learning workloads. The Poplar software framework provides the compiler, runtime, and libraries that make IPU programming accessible to PyTorch and TensorFlow users while exposing the architectural advantages of the IPU to programmers who want maximum performance. For AI researchers and enterprises seeking alternative hardware platforms that offer distinctive performance characteristics for specific workload profiles — particularly sparse models, graph neural networks, and models with irregular memory access patterns — Graphcore's IPU technology represents one of the most genuinely innovative approaches available in the AI hardware ecosystem today.