Meta Platforms picks up AI networking chip team from Graphcore

Meta Platforms Inc has recruited an Oslo-based team that until late last year was creating artificial-intelligence networking technologies at British semiconductor startup Graphcore.

A Meta representative verified the hirings in response to a request for comment, after Reuters found 10 persons whose LinkedIn accounts claimed they worked for Graphcore until December 2022 or January 2023 and later joined Meta in February or March of this year.

"We just welcomed a lot of highly skilled engineers in Oslo to our infrastructure team at Meta. They provide substantial knowledge in the design and development of supercomputing systems to enable AI and machine learning at scale in Meta's data centers," said Jon Carvill, the Meta spokesman.

The move offers more power to the social media giant's attempt to enhance how its data centers handle AI work, as it rushes to deal with demand for AI-oriented infrastructure from teams throughout the firm eager to create new products.

Meta, which owns Facebook and Instagram, has been more dependent on AI technology to target advertising, choose posts for its applications' feeds and remove forbidden material from its platforms.
On top of that, it is now hurrying to join rivals like Microsoft Corp and Alphabet Inc's Google in producing generative AI products capable of generating human-like writing, art and other material, which investors view as the next major growth sector for tech businesses.

The 10 individuals' job profiles on LinkedIn showed the team had worked on AI-specific networking technologies at Graphcore, which creates computer chips and systems tailored for AI work.
Carvill refused to specify what they would be working on at Meta.

Graphcore shuttered its Oslo office as part of a bigger restructure announced in October last year, a representative for the business said, as it struggled to make inroads against U.S.-based rivals like Nvidia Corp and Advanced Micro Devices Inc who dominate the market for AI processors.

Meta already has an in-house team building multiple sorts of chips geared at speeding up and increasing efficiency for its AI work, including a network chip that serves a kind of air traffic management role for servers, two individuals told Reuters.

Efficient networking is particularly helpful for current AI systems like those powering chatbot ChatGPT or image-generation tool Dall-E, which are much too huge to fit onto a single processor chip and must instead be divided out among multiple chips connected together.

A new sort of network chip has evolved to assist keep data traveling smoothly inside such computer clusters. Nvidia, AMD and Intel Corp all produce similar network processors.

In addition to its network chip, Meta is also creating a complicated computational chip to both train AI models and execute inference, a process in which the trained models make judgements and create replies to prompts, although it does not anticipate that chip to be available until around 2025.

Graphcore, one of the UK's most valuable tech businesses, formerly was viewed by investors like Microsoft and venture capital company Sequoia as a promising prospective contender to Nvidia's dominating dominance in the market for AI chip systems.

However, it encountered a setback in 2020 when Microsoft abandoned an early arrangement to acquire Graphcore's processors for its Azure cloud computing platform, according to a story by UK daily The Times. Microsoft instead employed Nvidia's GPUs to construct the huge infrastructure enabling ChatGPT developer OpenAI, which Microsoft also backs.

Sequoia has subsequently written down its stake in Graphcore to zero, but it remains on the company's board, according to a person familiar with the situation. The write-down was initially revealed by Insider in October.

The Graphcore representative admitted the difficulties, but claimed the business was "perfectly positioned" to take advantage of rising commercial usage of AI.
Graphcore was reportedly valued at $2.8 billion after receiving $222 million in its most recent funding round in 2020.