Meta announced the MTIA 400, a custom AI accelerator for large language model training and ad recommendation inference, at the Hot Chips conference.

This is Meta's first purpose‑built generative AI accelerator. The company previously created custom silicon for ad serving.

The MTIA 400 uses a heterogeneous multi‑die architecture with two compute chiplets on a 3 nm process. Each chiplet contains a 6 × 8 grid of processing elements delivering 12 petaFLOPS of MXFP4 at 1.7 GHz.

Memory is provided by eight 36 GB HBM3e stacks, for a total of 288 GB and 9.2 TB/s bandwidth.

Inter‑chip bandwidth is 1.2 TB/s over RDMA.

The chip is about 20 % faster than Nvidia’s top‑specification Blackwell chips at the precision levels used for model training, while using similar power.

It is three to three and a half times slower than Nvidia’s Rubin GPUs and AMD’s Instinct MI455X accelerators.

Each compute blade contains four MTIA 400 chips and connects to an x86 CPU and NIC via a PCIe switch.

A rack holds 18 blades and eight switch blades, providing 72 accelerators in a single domain.

Meta plans an MTIA 450 with doubled memory bandwidth using HBM4, and an MTIA 500 in 2027 with further bandwidth and compute growth.

Meta has not said the MTIA 400 will replace AMD or Nvidia GPUs, but expects it to support its expanding AI workloads and upcoming inference‑optimized recommender models.