SAN FRANCISCO, March 11 - Meta Platforms unveiled a multi-chip plan on Wednesday for processors the company is designing internally as it expands the number of data centers that run its social apps.
The chips are being developed through Meta's Meta Training and Inference Accelerator (MTIA) program. The initial model from the latest cohort - the MTIA 300 - is already serving production duties, powering the company's ranking and recommendation systems. Meta said three additional chips will follow, with deployments scheduled to occur this year and continuing into 2027.
Meta and other large technology firms have been building internal chip-design teams alongside purchases of commercially available GPUs from vendors such as Nvidia and Advanced Micro Devices. The company said the rationale for in-house designs is to produce processors tailored to the specific types of computation Meta needs, which can yield gains in energy efficiency and cost.
Two of the upcoming designs - identified as the MTIA 450 and MTIA 500 - are explicitly intended for inference, the stage when a trained AI model answers user queries and executes tasks. "We see inference demand exploding at the moment and that's what we're currently focused on," Yee Jiun Song, Meta's vice president of engineering, said in an interview.
Meta acknowledged progress with chips built for inference, noting some success, but said it has faced challenges pursuing a generative AI training chip suitable for constructing the very large models that underpin advanced AI applications. The company described a shift in system-level design beginning with the MTIA 400. That design effort encompasses an integrated system roughly the size of several server racks, and incorporates a version of liquid cooling.
The company plans to introduce new chips on roughly six-month cadences as it scales its infrastructure. "That is the reality of how quickly our infrastructure is being built out," Song said, linking the accelerated release schedule to the pace of data center expansion required to operate apps such as Instagram and Facebook.
Meta disclosed in January that it expects 2024 capital expenditures to fall between $115 billion and $135 billion. The company also continues to use external partners for aspects of chip production and design. Meta has contracted Broadcom to assist with certain elements of its chip designs, though Song did not specify which chips Broadcom is supporting. For fabrication, Meta relies on Taiwan Semiconductor Manufacturing Co to produce the processors.
In addition to its in-house initiative, Meta struck large purchase agreements in February with Nvidia and AMD to buy chips worth tens of billions of dollars. These external purchases sit alongside Meta's internal MTIA program as the company balances custom designs with third-party hardware commitments.
Context and implications
The four-chip roadmap under MTIA reflects Meta's dual approach to meet rapidly growing compute needs: continue large-scale procurement of externally manufactured GPUs while developing specialized accelerators optimized for the company's workloads. The MTIA 300's deployment to ranking and recommendation systems demonstrates a focus on production-ready work that drives user-facing services, while the later MTIA 450 and 500 target inference workloads where latency and efficiency are critical.
Meta's system-level work beginning with MTIA 400 - including rack-scale designs and liquid cooling - signals an integration of chip design with data center infrastructure to support higher density and thermal management demands.