The training chip, designed specifically for AI-related tasks, was reportedly manufactured by TSMC (Taiwan Semiconductor Manufacturing Company). The test deployment follows a successful tape-out, the final stage before semiconductor production. Meta has not yet commented on the development.
Meta’s Push for Self-Reliance in AI Hardware
Meta’s move to develop its own chips, known as the Meta Training and Inference Accelerator (MTIA), has been in the works for several years. First reported in 2023, the MTIA chips are based on 7nm nodes and offer 102 Tops of Integer (8-bit) accuracy computation or 51.2 teraflops of FP16 accuracy computation. The chips operate at 800 megahertz and measure approximately 370 millimeters square.
Originally slated for release in 2022, the project was delayed after the chips failed to meet internal performance targets. The shift from CPUs to GPUs for AI training also necessitated a redesign of Meta’s data centers, leading to the cancellation of several projects.
Second-Generation MTIA Chips and Strategic Acquisitions
In February 2024, Reuters reported that Meta was planning to deploy the second generation of its MTIA chips, signaling continued investment in its in-house AI hardware capabilities. Additionally, Meta has explored strategic acquisitions to bolster its chip development efforts, including talks to acquire South Korean AI chip startup FuriosaAI.
Implications for Meta’s AI Ambitions
The development of in-house AI training chips aligns with Meta’s broader strategy to enhance its AI infrastructure and reduce reliance on third-party hardware providers like Nvidia. By controlling its chip production, Meta aims to optimize performance, reduce costs, and gain greater flexibility in scaling its AI capabilities.
As Meta continues to test and refine its MTIA chips, the tech industry will be watching closely to see how this move impacts the competitive landscape of AI hardware and Meta’s ability to innovate in the rapidly evolving field of artificial intelligence.