by MetaUS
Design Meta's own AI chips for the ranking and recommendation models behind its feeds and ads. Running these huge daily workloads on custom silicon cuts cost and reliance on NVIDIA.
MTIA 300 in production; MTIA 400 deploying for recommendation and GenAI; MTIA 450, 500 in 2027.
Checked 11 Oct0 updates this week
MTIA is designed for ranking posts and generating recommendations, and increasingly for generative AI inference.
Custom silicon beats general chips. MTIA uses lower-precision arithmetic to cut power and memory bandwidth.
MTIA moved from first silicon to running in 16 regions in under nine months, scaling to tens of thousands of chips across Meta's data centres.
Meta releases a new generation every six months to keep pace with demand for both recommendation and generative AI inference workloads.
| Spec | MTIA |
|---|---|
| MTIA 300 process node | TSMC 5nmR (reported) |
| MTIA 300 memory | 256MB SRAM + 128GB LPDDR5R (reported) |
| MTIA 300 power | 90W TDPR (reported) |
| MTIA 300 peak performance | 708 TFLOPS (INT8)R (reported) |
| MTIA 300 speedup over v1 | 3× on evaluated modelsR (reported) |
| MTIA 300 platform throughput improvement | 6× over v1R (reported) |
R reported by the company
Meta
Llama open models, Muse, AI data centers
Meta, owner of Facebook, Instagram and WhatsApp, made open-weight AI mainstream with its Llama models from 2023. In 2025 it formed Meta Superintelligence Labs. In April 2026 it released Muse Spark, its first closed frontier model.
