Details
- Meta announced advancements in its MTIA chip family, deploying MTIA 300 for R&R training and preparing MTIA 400, 450, and 500 for GenAI workloads through 2027.
- Involves Meta's in-house development with Broadcom partnership; chips deployed in hundreds of thousands across data centers powering billions of users.
- New generations feature modular chiplets, doubled HBM bandwidth in MTIA 450 vs. 400, 50% more in MTIA 500, low-precision data types like MX4, and PyTorch-native software stack including vLLM support.
- Evolves from MTIA 100/200 (ISCA papers) with 4.5x HBM bandwidth and 25x compute FLOPS gains from 300 to 500; iterative 6-month cadence adapts to shifting AI models unlike slower GPU cycles.
- Search confirms MTIA production scaling with TSMC 3nm for later versions in H1/H2 2026, aligning with reports of MTIA-2/3 tape-outs and ASIC market growth to 27.8% by 2026.
Impact
Meta's high-velocity MTIA strategy reduces reliance on Nvidia GPUs for inference-heavy workloads, targeting cost-effective scaling for GenAI serving billions daily. With modular designs and PyTorch integration, it accelerates deployment amid ASIC surge from Google and others, potentially reshaping hyperscaler chip economics by 2027.
