Meta Platforms just lately unveiled its in-house custom chip “family” aimed toward enhancing artificial intelligence (AI) work. The firm developed its first-generation chip in 2020 as a part of the Meta Training and Inference Accelerator (MTIA) programme, to improve effectivity for recommendation fashions used in serving adverts and different content in information feeds.
The first MTIA chip was designed solely for an AI course of called inference, the place algorithms trained on massive quantities of information make judgments about which content material to show next in a user’s feed.
Joel Coburn, a software engineer at Meta, defined that the corporate initially used graphics processing units (GPUs) for inference tasks but found them ill-suited for the job. He stated…
“Their effectivity is low for real fashions, despite significant software program optimizations. This makes them difficult and costly to deploy in practice. This is why we want MTIA.”
A spokesperson for Meta didn’t present particulars on deployment timelines for the new chip or plans to develop chips for coaching models. The company has been working on upgrading its AI infrastructure in the past yr after recognising that it lacked the necessary hardware and software program to assist AI-powered options. Consequently, Meta scrapped plans for a large-scale rollout of its in-house inference chip and commenced engaged on a more bold chip able to performing each coaching and inference duties.
While the primary MTIA chip struggled with high-complexity AI fashions, it managed low- and medium-complexity models extra efficiently than competitor chips. The chip consumed solely 25 watts of power, considerably lower than market-leading chips from suppliers like Nvidia Corp, and used an open-source chip structure called RISC-V.
Controversial introduced plans to redesign its knowledge centres with trendy AI-oriented networking and cooling methods, with the primary facility set to interrupt ground this 12 months. The new design is anticipated to be 31% cheaper and built twice as fast as the company’s current data centres..