Meta is forging ahead with the development of its next-generation infrastructure, with a strong focus on integrating AI capabilities. This strategic move encompasses the support for new generative AI products, recommendation systems, and cutting-edge AI research. The company anticipates substantial growth in this area as the demand for compute power to sustain AI models rises in tandem with their complexity.

Last year, Meta introduced the Meta Training and Inference Accelerator (MTIA) v1, marking its inaugural foray into AI inference acceleration. Tailored specifically for Meta’s deep learning recommendation models, MTIA v1 aimed to enhance compute efficiency across Meta’s suite of apps and technologies.

MTIA represents a significant long-term investment geared towards providing an optimal architecture for Meta’s distinctive AI workloads. As AI continues to play a pivotal role in Meta’s products and services, this efficiency becomes paramount in delivering top-tier user experiences worldwide. The introduction of MTIA v1 marked a crucial milestone in bolstering compute efficiency within Meta’s infrastructure and empowering software developers to craft AI models that drive enhanced user interactions.

The next iteration of MTIA is part of Meta’s comprehensive full-stack development initiative for custom, domain-specific silicon catering to its unique workloads and systems. This upgraded version boasts over double the compute and memory bandwidth of its predecessor while maintaining close alignment with Meta’s workloads. Engineered to efficiently support ranking and recommendation models, this chip prioritizes the optimal balance of compute power, memory bandwidth, and capacity.

MTIA is poised to become a cornerstone of Meta’s long-term roadmap for building and scaling powerful and efficient infrastructure tailored to its AI workloads. Meta is strategically designing its custom silicon to seamlessly integrate with existing infrastructure while remaining adaptable to future advancements, including next-generation GPUs. Achieving Meta’s ambitions for custom silicon entails substantial investments not only in compute silicon but also in memory bandwidth, networking, capacity, and other next-generation hardware systems.

By Impact Lab