Meta AI Chip Targeted Recommenders

Meta has developed an AI accelerator chip called MTIA, following in the footsteps of Google and Amazon. Employing RISC V CPUs and fixed-function units, the chip offered 102 INT8 TOPS at 25 W TDP.

Anand Joshi

Seeking an AI processor that better met its needs than merchant-market chips, Facebook-parent Meta, developed its own. Development of the Meta Training and Inference Accelerator (MTIA) started in 2020, and the company planned a full-scale 2022 rollout. By then it proved less viable, and the company decided not to widely use it.

Meta, however, recently disclosed MTIA details, and reports indicate that it’s working on a successor. The defunct MTIA targeted the deep-learning recommendation model (DLRM) versions that make up most of Meta’s AI workload. An important aim was to deliver better performance per watt than commercial alternatives to reduce total cost of ownership (TCO).

Requiring only 25W, the MTIA offered raw performance of 102 INT8 TOPS, or 4 TOPS per watt—competitive with the era’s chips. This low power enabled the company to fit it into a small, power-constrained card to be added to its servers.

Meta’s data show MTIA to be mediocre, performing well on smaller DLRM versions compared with alternatives but worse on larger models. The company has already started working on a successor, targeting a 2025 release. Stable software, competitive performance per watt, and the flexibility to run new models will determine the success of the program.