- According to the reports, Meta is preparing its first internal AI training chip for deployment
- The dedicated AI accelerator, made with TSMC, completed the adhesive tape
- The goal change to personalized silicon aims to reduce its dependence on the NVIDIA hardware
Like many of Nvidia’s highest customers, Meta is looking to reduce their dependence on the expensive Hardware of the GPU manufacturer by making their own silicon.
In 2024, the social networks giant began to announce engineers to help build their own latest generation automatic learning accelerators, and now, according to an exclusive report from PakGazetteGoal is in the trial stage for its first internal chip designed to train AI systems.
Sources said PakGazette That after its first chip adhesive tape, a target has begun a limited implementation, and if the tests are going well, it plans to climb the production for broader use.
RISC-V business
According to PakGazette, “the new goal training chip is a dedicated accelerator, which means that it is designed to handle only specific tasks of AI. This can make it more efficient energy than the integrated graphics processing units (GPU) generally used for AI workloads.”
The chip manufacturer based in Taiwan, TSMC, produced the silicon for the finish line as part of the training and inference accelerator program of the Facebook owner (Mtia), something that PakGazette Signs has had “a staggering beginning for years and at a time rejected a chip in a similar phase of development.”
In 2023, Meta released its first -generation inference accelerator of AI designed to feed the classification and recommendation systems for Facebook and Instagram, and then, in April 2024, a new version debuted that doubled the bandwidth of memory and memory.
In the Hot Hot 2024 Chips Symposium, Meta revealed that its inference chip was based on the 5 NM TSMC process, with the processing elements in the RISC-V nuclei.
Like a growing number of technology companies, Facebook has delayed its weight to recognize its AI ambitions, and although the PakGazette The report does not provide any detail about the technical aspects of the new training chip of the Meta, it seems a fair commitment that will also be based on open source RISC-V architecture.
He PakGazette The article indicates that executive goal say they want to start using their own chips to train for next year.