Fractile

Fractile is a London-based semiconductor company positioning itself to address a critical bottleneck in the artificial intelligence sector: inference performance. Founded in 2022 by CEO Dr. Walter Goodwin and co-founder Yuhang Song, the company operated in stealth mode for two years before emerging in 2024. Goodwin's background includes a Ph.D. in AI and robotics from the University of Oxford, where his research on using large foundation models for general-purpose robotics revealed the hardware limitations that inspired Fractile's creation.

The company is developing a new class of AI accelerator chips specifically designed to run large language models (LLMs) during the inference stage—the process of using a trained model to generate outputs. Current hardware, primarily GPUs, is optimized for training models but is inefficient for inference, where the constant movement of a network's weights onto a chip creates a significant performance lag. Fractile's core value proposition lies in its in-memory computing architecture, which fuses computation directly with memory. This design circumvents the data transfer bottleneck, projecting performance improvements of up to 100 times faster execution at one-tenth of the system cost compared to current market-leading hardware.

Fractile's business model centers on designing and supplying these specialized chips and the accompanying full software stack to clients who deploy large-scale AI models, such as cloud providers and large AI companies. By enabling significantly faster and cheaper inference, the company aims to unlock new capabilities and broader adoption of advanced AI in computationally heavy fields like drug discovery and climate modeling. The company has secured substantial financial backing, raising a total of $17.5 million from investors including Kindred Capital, Oxford Science Enterprises, and the NATO Innovation Fund. Noteworthy angel investors include former Intel CEO Pat Gelsinger, Hermann Hauser, and Stan Boland, signaling strong confidence from semiconductor industry veterans.

Keywords: AI accelerator, semiconductor, in-memory computing, large language model inference, AI hardware, data center accelerator, chip design, LLM inference, AI chip, GPU alternative, deep learning hardware, neural network processor, robotics AI, high-performance computing, memory bottleneck, low-latency AI, efficient AI, sequence processing, AI infrastructure, model deployment, CMOS SRAM, Pat Gelsinger, Walter Goodwin, Oxford Science Enterprises, NATO Innovation Fund