Quick Facts｜Meta Announces Latest Custom ai Chip Progress, Expects $18 Billion in GPU Investment by 2024

Article source:There's a new Newin

Meta just announced an update on its custom AI chip.Meta debuted its first-generation AI inference gas pedal last year, the Meta Training and Inference Accelerator (MTIA) v1, designed specifically for Meta's AI workloads. It is particularly suited for deep learning recommendation models that improve a variety of experiences across its applications and technologies.

MTIA is designed to provide the most efficient architecture for Meta's unique workloads. As AI workloads become increasingly important in their products and services, this efficiency is key to Meta's ability to deliver the best possible experience to its users around the world. MTIA v1 is an important step in improving infrastructure compute efficiency and better supporting software developers in building new AI models that improve the user experience.

Meta's next-generation, large-scale infrastructure is being built with AI needs in mind, including support for new generative AI products, recommender systems, and advanced AI research. Investments in this area are expected to increase in the coming years as the computational requirements to support AI models and the complexity of the models grow.

The next-generation MTIA is part of Meta's full-stack development program for custom, domain-specific chips designed to address its unique workloads and systems. This new version of MTIA more than doubles the compute and memory bandwidth of previous solutions while maintaining a tight connection to the workload.

It is designed to efficiently serve ranking and recommendation models that provide high quality recommendations. The architecture of this chip focuses on providing the right balance between computation, memory bandwidth and memory capacity for ranking and recommendation models.

MTIA has been deployed in Meta's data centers and is being served in production environments. Through this program, Meta is able to dedicate more compute power to more intensive AI workloads.

Results to date show that this MTIA chip is capable of handling both low and high complexity ranking and recommendation models, which are a key part of the Meta product. Because Meta controls the entire technology stack, it is able to achieve greater efficiency than commercially available GPUs.

Meta is continuing to invest in custom silicon. mTIA will be an important part of Meta's long-term roadmap to build and scale the most powerful and efficient infrastructure for its unique AI workloads.

Meta is designing its custom chips to work with existing infrastructure as well as new, more advanced hardware that may be utilized in the future, including next-generation GPUs. Achieving the goal of customized chips means investing not only in compute chips, but also in memory bandwidth, networking and capacity, and other next-generation hardware systems.

Currently, Meta is working on several projects aimed at expanding the application of MTIA to include support for generative AI workloads. This journey has just begun.

PS: Looking back at 2023 so far, most of the AI hype has focused on the horizontal capabilities of the underlying models, but the real opportunity for AI lies in the How AI and Agents are reconfiguring and creating the B2B value chainThe

Reference:

https://about.fb.com/news/2024/04/introducing-our-next-generation-infrastructure-for-ai/