Meta is developing its first in-house AI training chip. According to Reuters, the company has begun pilot testing a small batch of the new AI training chip and will order more if it succeeds in the pilot test. The chip is being manufactured by TSMC.
Meta's decision to build its own AI chip is also part of a longer-term strategy to reduce the company's reliance on third-party suppliers like Nvidia, which is a dominant provider of graphics processing units (GPUs) used in AI-related workloads. The tech giant has projected its overall expenses, including capital expenditures, to reach as high as $119 billion by 2025—much of it to create AI infrastructure.
The Meta-designed chip is a single-use AI accelerator. It is specialized for AI-related tasks compared to general computation. Such specialized architecture can render the chip more power-efficient compared to the GPUs currently used in AI training.
Meta first plans to use the chip in its recommendation algorithms that determine content displayed on Facebook and Instagram. The company eventually aims to scale up the use of the chip to power its generative AI products, such as the AI chatbot known as Meta AI.
Meta's foray into custom silicon development has had a mixed track record so far. The company previously scrapped an earlier in-house inference chip after a failed small-scale deployment, opting instead to purchase billions of dollars worth of Nvidia GPUs. However, the new training chip appears to be a more successful endeavor, as Meta has passed the crucial "tape-out" milestone in its development.
Meta is not the only company developing its own AI chip. In other news, OpenAI is completing its first custom AI training chip design. The new chip will likely feature a systolic array architecture along with high-bandwidth memory, like Nvidia's latest AI accelerators. This architecture is known for its high performance and efficiency in processing dense computations.
Source: Reuters
4 Comments - Add comment