Meta rolls out MTIA 300 and 400 custom AI chips to reduce Nvidia reliance

Meta

Meta announced its MTIA custom AI accelerator roadmap, deploying MTIA 300 in production for recommendation models and launching MTIA 400 for generative AI inference. The chips reduce reliance on off-the-shelf processors and cut costs. Meta plans releases every six months through MTIA 500 in 2027.

• MTIA 300 offers 3× speedup on recommendation models and 6× throughput improvement over the original MTIA v1.
• MTIA 400 targets generative AI inference with support for lower-precision operations (FP4) and scales to 72-node domains.
• Six-month cadence lets Meta match rapid advances in AI model sizes and demands.
• Custom silicon reduces power consumption and memory bandwidth compared to general-purpose GPUs.