Photo credit: www.techradar.com
Meta is on the verge of launching its first proprietary AI training chipCollaboration with TSMC marks a significant milestoneMeta aims to decrease reliance on Nvidia’s hardware
In a move reflective of wider industry trends, Meta is actively pursuing the development of its own AI chips, aiming to lessen dependence on the costly GPUs provided by Nvidia.
Starting in 2024, the tech giant began recruiting engineers focused on the creation of advanced machine learning accelerators. Recent reports from Reuters indicate that Meta has progressed to the testing phase for its inaugural in-house chip specifically tailored for AI training.
According to sources cited by Reuters, following the initial tape-out of the chip, Meta has launched a limited deployment. Should testing prove successful, the company envisions scaling up production for broader applications.
The RISC-V Initiative
As per the analysis by Reuters, Meta’s new training chip is classified as a dedicated accelerator, designed specifically for AI-related tasks. This specialization can potentially lead to improved power efficiency compared to traditional GPUs which are often employed for AI functions.
Meta has partnered with Taiwan Semiconductor Manufacturing Company (TSMC) for the production of these chips, as part of its Meta Training and Inference Accelerator (MTIA) initiative. This program has experienced challenges in its early stages, including the abandonment of a chip during an analogous phase of development.
In 2023, Meta introduced its first-generation in-house AI inference accelerator aimed at enhancing the ranking and recommendation systems used on platforms like Facebook and Instagram. By April 2024, the company unveiled an improved version that features twice the compute power and memory bandwidth of its predecessor.
During the 2024 Hot Chips symposium, Meta disclosed that the inference chip was manufactured using TSMC’s advanced 5nm process, with its processing components built around RISC-V architecture.
Aligning with a growing trend in the tech industry, Meta is embracing the RISC-V open-source architecture to further its AI goals. While the technical specifics of the new AI training chip remain largely undisclosed, it is probable that it will also utilize RISC-V frameworks, aligning with Meta’s strategic objectives.
Executives at Meta have expressed aspirations to begin employing their custom chips for AI training as early as next year.
Related Topics You May Find Interesting
Source
www.techradar.com