MTIA (Meta Training and Inference Accelerator) is Meta's in-house AI chip family. Design began in 2020; MTIA v1 (7 nm) was announced in May 2023 and v2 (5 nm)…
No. Its first public AI hardware, Big Sur in 2015, was a server built around eight Nvidia GPUs.
Meta began designing its first MTIA chip in 2020, for the recommendation models behind feeds and ads.
Yes. Reuters reported an in-house inference chip underperformed in 2022, and Meta bought billions of dollars of Nvidia GPUs instead.
MTIA v1, now MTIA 100: a 7 nm, 25-watt chip for recommendation inference, announced on 18 May 2023.
MTIA 200 moved to 5 nm and showed three times v1's performance on key models, already serving 16 data-centre regions.
No. In February 2026 Meta ordered millions of Nvidia GPUs and 6 gigawatts of AMD GPUs.
The Information reported Meta dropped Olympus, its most advanced training chip, after design problems.
Four generations in about two years: 300 for recommendation training, 400 to 500 increasingly for GenAI inference.
Serving language models is limited by memory bandwidth, so MTIA 450 doubles HBM bandwidth over 400.
Broadcom co-designs MTIA, its packaging and networking, under a deal to 2029 starting above 1 gigawatt.
PyTorch, Triton and vLLM support, plus KernelEvolve, an AI agent that writes faster kernels.
Eight HBM3e stacks and about 12 petaflops of FP4; Meta calls it competitive with leading commercial chips.
A Broadcom-designed, TSMC-made MTIA chip that Reuters reported would enter production in September 2026.
MTIA 450 mass deployment in early 2027, MTIA 500 later in 2027, and whether the six-month cadence holds.
Read the complete story, FAQs and key facts.