Meta is to put the third of its multi-generational MTIA (Meta Training and Inference Accelerators) into production in September as part of its plan to have 14GW of computing power in place in 2027.
The chip – dubbed MTIA v3 and nicknamed Iris (pictured) – was designed with Broadcom and is for handling both heavy AI training models and low-latency inference workloads. It will be fabbed on TSMC’s 3nm process.
MITA
Although the five year-old MTIA programme has slipped from its original targets, the programme now intends to launch a new chip generation every six months: MTIA v4 (Santa Barbara) in 2027, v5 (Olympus) also in 2027, and v6 (Universal Core) in 2028.
Meta reckons that using its own chips in its datacentres delivers a 40% to 44% reduction in TCO.
Meta is in catch-up mode with its Mag7 peers in using its own custom chips. Google started using its own TPUs in its datacentres in 2014.
Last week Meta CEO Mark Zuckerberg said the company might rent out datacentre capacity.
See all our Datacentre content.