London-based Oriole Networks, which says it has developed novel technology to help train AI faster, has raised £17.5m in a Series A round, according to Companies House filings.
Ian Hogarth, partner at London-based VC firm Plural, was appointed as a director of the company the same day in August as new shares were issued, indicating that the investor was involved in the round.
Oriole Networks and Plural did not respond to a request for comment.
The startup is developing a new way to connect up graphic processing units (GPUs) — the chips used to train large generative AI models — using photonics, a technology that uses light to transmit computing information.
Oriole says that its tech can increase the speed that information can travel by up to 100 times, alongside cutting energy consumption to 2-3% of the amount used by ethernet cables — the way information is traditionally transferred between GPUs.
The company’s Series A comes just six months on from its £10m seed round. Backers in that round included XTX Ventures, the venture arm of algorithmic trading firm XTX, Clean Growth Fund, Dorilton Ventures and the UCL Technology Fund.
The future of training AI models?
Training AI system large language models (LLMs) — the tech that powers chatbots like ChatGPT — means clustering thousands of GPUs together, usually with ethernet cables.
But as GPUs have become significantly more advanced over the past decade, the ethernet networks that connect them haven’t, CEO James Regan told Sifted in March, resulting in a “bottleneck”.
Ethernet networks can slow down the compute power of the cluster when they hit a certain amount of information passing through them. They also consume a significant amount of energy — roughly 20% of the total consumption of the overall AI cluster (the rest is from computer processing itself).
Oriole’s not the only company to spot an opportunity; Google is also building an optical network for AI training, in a project called Mission Apollo.
The startup spun out from University College London’s tech transfer arm in 2023 and owns the IP for the physical architecture of its optical network, alongside the machine learning technology that makes it possible.
Regan told Sifted in March that his plan to commercialise involved selling the networking system to people building data centres and other compute infrastructure, and to outsource the actual hardware manufacturing of the product.
Oriole hasn’t yet released a product and is pre-revenue, but Regan said that the company hopes to take its technology to market in a couple of years.
Read the orginal article: https://sifted.eu/articles/oriole-networks-17m-raise-plural-news/
