Meta Unveils Next-Generation AI Chip

While consuming more power, it delivers up to three times better performance compared to its predecessor.

Meta Unveils Next-Generation AI Chip
Photo by Dima Solomin / Unsplash

Meta, has unveiled the latest iteration of its Meta Training and Inference Accelerator (MTIA), marking an advancement in its custom-made chip technology. The MTIA v2 promises three times the performance of its predecessor, MTIA v1, while maintaining a strong alignment with Meta's specific AI workloads, boasting a larger design with more processing cores, increased internal memory, and higher clock speeds.

Notably, Meta is deploying the next-gen MTIA in its data centers for tasks such as ranking and recommending display ads on its platforms like Facebook. However, the company has acknowledged that it is not currently using the chip for generative AI training workloads, although it is exploring various programs in this direction. Instead, the next-gen MTIA is intended to complement GPUs rather than replace them for running or training models.

To support the next-generation silicon, Meta has developed a large, rack-based system capable of accommodating up to 72 accelerators. Running at 1.35GHz and consuming 90 watts of power, the MTIA v2 system offers denser capabilities and higher performance compared to the previous design, ensuring efficient operation in Meta's data centers.

Meta has prioritized software integration, ensuring seamless compatibility with PyTorch 2.0 and optimizing the MTIA stack with features like TorchDynamo and TorchInductor. The Triton-MTIA compiler backend enhances developer productivity by generating high-performance code tailored for MTIA hardware.

MTIA v2 is already deployed in Meta's data centers, actively serving production models, and working in tandem with commercially available GPUs to deliver optimal performance and efficiency for Meta's specific AI tasks. As Meta continues its AI innovation journey, MTIA will evolve to support future endeavors, including generative AI workloads and enhancements in memory bandwidth, networking, and capacity, aligning with Meta's long-term AI goals.