Meta pilots first in-house AI training chip, production already underway
Reuters reports that Meta, which develops large-scale language models such as Llama, has experimentally introduced its first in-house chips for AI training. NVIDIA has a large share of the chip market for AI training, but introducing its own chips is an important step for Meta to reduce its reliance on NVIDIA.
Exclusive: Meta begins testing its first in-house AI training chip | Reuters
Meta is reportedly testing its first RISC-V based AI chip for AI training | Tom's Hardware
https://www.tomshardware.com/tech-industry/artificial-intelligence/meta-is-reportedly-testing-its-first-rsic-v-based-ai-chip-for-ai-training
Meta has been developing its own AI chips, including 'MTIA ,' for some time now, and has reportedly been planning to introduce its own custom chips into its own data centers from the second half of 2024.
Meta to deploy proprietary AI processor 'Artemis' in its data centers in the second half of 2024 - GIGAZINE
According to a source who spoke to Reuters, the accelerator being introduced this time is an AI-specific accelerator that Meta developed in-house. The chip is optimized for AI tasks, so it may be more power-efficient than using a GPU. Meta is trialing its own chips, and if the tests go well, it plans to expand production of the chips for large-scale use, Reuters reports.
The trial deployment began after the first 'tape-out' was completed, according to the people familiar with the matter, which means that the chip design information has been sent to TSMC's manufacturing plants, stepping up from chip design to manufacturing.
Tape-outs typically cost tens of millions of dollars and take about three to six months to complete, but there's still no guarantee that the test will work. If it fails, Meta will have to redesign and re-do the tape-out.
Tom's Hardware, a hardware news site, pointed out that this could be the industry's first AI accelerator developed based on the RISC-V architecture under a joint collaboration between Meta, Broadcom and TSMC. 'Meta's accelerators must demonstrate a level of power efficiency that competes with NVIDIA's H200 , B200 , and next-generation B300,' he said, arguing that meeting this goal would significantly reduce dependence on NVIDIA chips.
Reuters reached out to Meta and TSMC for comment, but both companies declined to comment.
Related Posts:
in Hardware, Posted by log1i_yk