Meta Unveils In-House AI Chips
Meta has unveiled four proprietary AI chips (MTIA) to optimize recommendation engines and generative AI workloads reported, reducing reliance on Nvidia and AMD.
Meta plans to deploy four generations of these AI chips by 2027: MTIA 300, 400, 450, and 500. MTIA 300 is already in production for ranking and recommendation training. MTIA 400 has completed testing and is slated for data center deployment soon. The MTIA 450 and 500 are scheduled for mass deployment in early 2027. Anticipating the rise in GenAI inference demand, MTIA 400 transitioned into MTIA 450, with specific optimizations for GenAI inference. The MTIA 500 will deliver 30 petaflops of MX4 inference performance and boasts 384 GB to 512 GB of HBM. Meta's iterative approach allows for chip designs to quickly adapt to evolving AI model requirements. The company is releasing new generations roughly every six months using a modular chiplet architecture. This enables faster adoption of new technology and keeps hardware aligned with changing models. These chips are designed to handle tasks from ranking and recommendation to general GenAI workloads. Meta's chips are application-specific integrated circuits (ASICs), designed for specific uses, unlike general purpose GPUs from Nvidia and AMD. By tailoring chip designs to their workload profiles, Meta aims to optimize performance and reduce costs.
