Meta poaches AI chip team from British tech unicorn

According to media reports, Meta poached a team from British artificial intelligence (AI) chip company Graphcore. The team previously worked in Oslo, Norway, and until the end of last year was still developing AI network technology at Graphcore.

It is reported that there are 10 people in the team. They worked at Graphcore until December last year or January this year, and then joined Meta in February or March this year.

In response to media requests for comment, Meta spokesman Jon Carvill confirmed that the company had indeed recruited the above-mentioned team.

Carvill said: “We have recently welcomed a number of highly specialized engineers to Meta’s infrastructure team in Oslo. They bring deep expertise in the design and development of supercomputer systems to support AI and AI at scale at Meta’s data centers. Machine Learning.”

Carvill declined to say what specific jobs the employees would do at Meta.

The job descriptions of these 10 employees on LinkedIn show that they were previously engaged in research on AI-specific network technology at Graphcore. Graphcore is one of the most valuable technology startups in the UK. It has been regarded by investors such as Microsoft and venture capital firm Redshirt Capital as having the potential to challenge Nvidia’s dominance in the AI ​​chip system market.

However, Graphcore suffered a setback in 2020, when Microsoft canceled an early deal to buy Graphcore chips for its Azure cloud computing platform and instead used Nvidia's GPUs to build the massive infrastructure that powers OpenAI, the company behind the chatbot ChatGPT.

In October last year, media reported that Sequoia Capital had written down its investment in Graphcore to zero. Graphcore announced restructuring plans at the time, including closing the company's Oslo offices.

Since its release in November last year, ChatGPT has set off a technology boom, with almost all technology giants participating in the field of generative AI, and Meta currently lags behind competitors such as Microsoft and Google in this field.

Sources say Meta has designed several chips internally to speed up its AI work and maximize efficiency, including a network chip that performs air traffic control functions for servers.

Efficient networks are particularly important for modern AI systems, such as those behind ChatGPT and the image generation tool Dall-E, which cannot be packed into a single computing chip and must be dispersed across many chips connected in series.

In addition to network chips, Meta is also designing a complex computing chip that can both train AI models and perform inference, but it is not expected to be ready until around 2025.