Fractile, the latest NVIDIA contender, emerged from stealth with $15M in seed funding
Fractile is the most recent startup to enter the AI chipmaking competition. The startup plans to develop chips that combine memory and processing into a single component, drastically boosting performance while slashing costs and energy consumption.
Founded in 2022 by the University of Oxford’s Robotics Institute graduate Walter Goodwin, Fractile is a startup working on chips for AI workloads that combine memory and processing into a single component, thus boosting the chips' performance while driving costs and energy consumption down. With its novel approach to chip design and manufacturing, Fractile is not only taking on NVIDIA but also startups working to get their share of the AI chip market, including Groq, Cerebras, Graphcore, and tech firms that have developed in-house AI chips, like Microsoft, Amazon, and Google.
GPUs became the standard for AI workloads because of their parallel processing capabilities (as opposed to CPUs, which can only linearly process data). However, GPUs were never meant to process the amounts of data neural networks require. GPUs rely on information stored in memory, which is a different component altogether. Moving data back and forth from the memory module to the processing unit is not an issue when dealing with graphics-intensive applications. Still, it has proven to be a bottleneck limiting GPU performance in AI workloads. By eliminating the distance, Fractile's simulations predict that the startup's chips could run an LLM about 100x faster and 10x more affordably than NVIDIA's offerings.
Fractile is aware it still has a long way to go. In particular, it is aware that the NVIDIA CUDA system has a solid developer base, and that other companies have tried to migrate developers from CUDA to their software systems with mixed results. Fractile is working on a proprietary software stack that should be lighter and simpler, as it does not need some of CUDA's components whose sole purpose is to prepare the GPUs to run workloads they were not designed to run originally. Fractile has no reliable guarantee that developers will be flocking to migrate to its software stack. However, this may be a concern it can put temporarily on hold, at least until the company starts its test chips production.
Fractile's $15 million seed round was co-led by Kindred Capital, the NATO innovation fund, and Oxford Sciences Enterprises, with participation from Inovia Capital and Cocoa. In addition to growing its team from a 14 to an 18 headcount, Fractile plans to spend its raised funds on further simulation testing, and later, on test chip production.
