Meta has unveiled its MTIA 400, a custom accelerator designed both to train large language models and to run the advertising recommendation systems that underpin its revenue. The combination is unusual because training models requires vast computing power, while ad recommendations are mainly limited by memory performance, but it could help Meta make more use of costly in-house hardware.

The chip uses a multi-die design with two 3nm compute dies, two I/O dies and a system-on-chip component, apparently drawing on Broadcom technology. Meta says it can deliver 12 petaFLOPS of MXFP4 performance at 1.7GHz, around 20% faster than Nvidia’s highest-specification Blackwell parts at certain training precisions with similar power use; however, it remains roughly three to 3.3 times slower than newer Nvidia Rubin and AMD MI455X accelerators, so is unlikely to replace their GPUs for frontier AI work soon.