Meta Rolls Out In-House AI Chips

Meta's MTIA chips are designed to handle AI workloads more efficiently, specifically improving the performance of recommendation systems. The first-generation MTIA chip was built using TSMC's 7nm process and ran at 800 MHz. While it showed only marginal improvements for simpler tasks, Meta planned to optimize the software to match GPU performance for complex tasks. Meta is rolling out four custom AI processors, including the MTIA 300 (already in production) and upcoming MTIA 400, 450, and 500 models, with new versions released every six months. The MTIA 400 is designed to match the performance of leading commercial AI chips, while enabling denser server configurations to lower operational costs. These chips target AI ranking and recommendation workloads. The latest MTIA chips show significant performance gains. Meta claims a 4.5x increase in memory bandwidth and a 25x jump in compute FLOPS from the MTIA 300 to the MTIA 500. The MTIA 400 delivers 400% higher FP8 FLOPS and 51% higher HBM bandwidth compared to the MTIA 300. Meta's move to develop in-house chips aims to reduce its reliance on external suppliers like Nvidia, and control costs. By tailoring chip designs to their specific workload profiles, Meta can optimize for inference, potentially reducing inference costs by 30-50% compared to commercial GPUs. This strategy mirrors similar moves by Google and Amazon.