- Meta would have prepared her first internal AI training chip for deployment
- The accelerator A dedicated, made with TSMC, completed by band
- The transition from meta to personalized silicon aims to reduce its dependence on Nvidia equipment
Like a lot of NvidiaThe highest spending customers, Meta seeks to reduce his dependence on the expensive IA equipment from the GPU manufacturer by doing their own silicon.
In 2024, the social media giant began Advertising for engineers To help build its own cutting -edge automatic learning accelerators, and now, according to an exclusive report ReutersMeta is at the test stage for her first internal chip designed for the training of AI systems.
Sources told Reuters This, after his first gang of the chip, Meta began a limited deployment, and if the tests go well, he plans to extend production for wider use.
RISC-V Business
According to Reuters, “Meta’s new training chip is a dedicated accelerator, which means that it is designed to manage only the specific tasks for AI. This can make it more powerful in power than integrated graphic processing units (GPU) generally used for workloads of the AI. »»
The TSMC flea manufacturer, based in Taiwan, produced silicon for Meta as part of the Facebook owner training and accelerator (MTIA), something that Reuters Soulies experienced “a wobbly start for years and at some point has abandoned a chip for a similar development phase”.
In 2023, Meta unveiled its first generation IA inference accelerator in the first generation designed to feed the classification and recommendation systems for Facebook and Instagram, then in April 2024, he made his debut a new version that doubled the calculation and memory band.
At the Hot Chips 2024 symposium, Meta revealed that her inference chip was built on the 5 nm of TSMC processWith treatment elements on Risc-V nuclei.
As an increasing number of technological companies, Facebook threw its weight behind Risc-V In order to recognize his ambitions of AI, and although the Reuters The report does not provide any details on the technical aspects of the new Meta AI training chip, it seems that it is just that it will also be based on Open Source RISC-V architecture.
THE Reuters The article notes that Meta leaders say they want to start using their own fleas for training by next year.