MTIA targets inference

Meta disclosed MTIA v1 on 18 May, a chip designed in 2020 for recommendation-model inference. Despite the programme’s Training and Inference name, this first accelerator runs trained models.

Photo: LPS.1, “Meta Platforms Headquarters Menlo Park California” (2022). CC0 1.0. Resized; cropped; responsive crops.

Its 25-watt thermal design power describes the chip. Meta separately reports boards consuming as little as 35 watts, so those figures cover different hardware boundaries.

Meta’s five-model tests measure performance per watt. It reports gains on low- and medium-complexity workloads but says high-complexity models remain unoptimised.

Reuters reports that Meta declined to give a deployment timetable and wasn’t planning a broad rollout of this first chip. The disclosure doesn’t establish production availability.

Alongside MTIA, Meta describes future liquid-cooled data centres and its 16,000-GPU Research SuperCluster. The projects serve different stages of its AI infrastructure.