Meta Platforms has employed an Oslo-based workforce that till late final yr was constructing artificial-intelligence networking expertise at British chip unicorn Graphcore.
A Meta spokesperson confirmed the hirings in response to a request for remark, after Reuters recognized 10 individuals whose LinkedIn profiles mentioned they labored at Graphcore till December 2022 or January 2023 and subsequently joined Meta in February or March of this yr.
“We lately welcomed various highly-specialized engineers in Oslo to our infrastructure workforce at Meta. They convey deep experience within the design and growth of supercomputing methods to assist AI and machine studying at scale in Meta’s knowledge facilities,” mentioned Jon Carvill, the meta spokesperson.
The transfer brings further muscle to the social media big’s bid to enhance how its knowledge facilities deal with AI work, because it races to deal with demand for AI-oriented infrastructure from groups throughout the corporate seeking to construct new options.
Meta, which owns Fb and Instagram, has turn into more and more reliant on AI expertise to focus on promoting, choose posts for its apps’ feeds and purge banned content material from its platforms.
On high of that, it’s now dashing to affix opponents like Microsoft and Alphabet’s Google in releasing generative AI merchandise able to creating human-like writing, artwork and different content material, which traders see as the following large development space for tech corporations.
The ten staff’ job descriptions on LinkedIn indicated the workforce had labored on AI-specific networking expertise at Graphcore, which develops pc chips and methods optimized for AI work.
Carvill declined to say what they’d be engaged on at Meta.
Graphcore closed its Oslo workplace as a part of a broader restructuring introduced in October final yr, a spokesperson for the startup mentioned, because it struggled to make inroads in opposition to US-based corporations like Nvidia and Superior Micro Units which dominate the marketplace for AI chips.
Meta already has an in-house unit designing a number of sorts of chips geared toward dashing up and maximizing effectivity for its AI work, together with a community chip that performs a kind of air site visitors management operate for servers, two sources instructed Reuters.
Environment friendly networking is particularly helpful for contemporary AI methods like these behind chatbot ChatGPT or image-generation software Dall-E, that are far too massive to suit onto a single computing chip and should as an alternative be cut up up over many chips strung collectively.
A brand new class of community chip has emerged to assist preserve knowledge transferring easily inside these computing clusters. Nvidia, AMD and Intel all make such community chips.
Along with its community chip, Meta can be designing a posh computing chip to each prepare AI fashions and carry out inference, a course of through which the educated fashions make judgments and generate responses to prompts, though it doesn’t anticipate that chip to be prepared till round 2025.
Graphcore, one of many UK’s most respected tech startups, was as soon as seen by traders like Microsoft and enterprise capital agency Sequoia as a promising potential challenger to Nvidia’s commanding lead available in the market for AI chip methods.
Nevertheless, it confronted a setback in 2020 when Microsoft scrapped an early deal to purchase Graphcore’s chips for its Azure cloud computing platform, in response to a report by UK newspaper The Instances. Microsoft as an alternative used Nvidia’s GPUs to construct the large infrastructure powering ChatGPT developer OpenAI, which Microsoft additionally backs.
Sequoia has since written down its funding in Graphcore to zero, though it stays on the corporate’s board, in response to a supply aware of the connection. The write-down was first reported by Insider in October.
The Graphcore spokesperson confirmed the setbacks, however mentioned the corporate was “completely positioned” to reap the benefits of accelerating business adoption of AI.
© Thomson Reuters 2023