Meta Unveils MTIA 400 AI Chip to Power LLM Training and Ad Serving
The fourth-generation custom accelerator blends generative AI training with recommender model workloads to reduce vendor reliance.
Meta has detailed its fourth-generation custom AI accelerator, the MTIA 400, marking a strategic shift toward supporting large language model (LLM) training. The announcement, made at the Hot Chips conference, signals the company's intent to vertically integrate its AI stack to manage the escalating costs of frontier model development.
According to reports from The Register, the MTIA 400 utilizes a heterogeneous multi-die architecture. The chip is built on a 3nm process and consists of two compute dies, two I/O dies, and a dedicated SoC die. In terms of raw power, the accelerator delivers 12 petaFLOPS of MXFP4 compute operating at 1.7 GHz. To handle massive datasets, Meta equipped the chip with eight 36 GB HBM3e stacks, totaling 288 GB of memory with a bandwidth of 9.2 TB/s. The hardware is deployed in racks containing 72 accelerators distributed across 18 compute blades, with each blade housing four MTIA 400 chips. Development of the silicon likely leveraged Broadcom's XPU technology.
A Pivot Toward Generative AI
Historically, Meta's custom silicon efforts focused on the specific, memory-bound requirements of deep learning recommender models (DLRM) used for ad serving. However, as the company pivots toward massive generative AI investments, the MTIA (Meta Training and Inference Accelerator) line has evolved. While the MTIA 400 maintains the ability to run ad-serving workloads, it is now designed primarily for the compute-intensive demands of LLM training. This dual-purpose design allows Meta to optimize its capital expenditure by using a single architecture for two of its most critical AI functions.
The Performance Gap
Despite its versatility, the MTIA 400 occupies a complex position in the competitive landscape. The Register reports that the chip is approximately 20% faster than Nvidia's top-specced Blackwell accelerators when utilizing the high precisions required for training. However, this lead is short-lived when compared to the next generation of hardware. The MTIA 400 is estimated to be 3x to 3.3x slower than Nvidia's Rubin and the Instinct MI455X, respectively.
This performance delta suggests that while custom silicon is a viable path for optimizing internal, specific workloads, it is not yet a wholesale replacement for the general-purpose power of frontier GPUs from Nvidia and AMD. Meta's strategy appears to be one of strategic diversification rather than total independence.
Future Roadmap
Meta is already planning the next iterations of its silicon to close these gaps. The company's roadmap includes the MTIA 450, which will be optimized for inference and feature HBM4 memory, as well as the MTIA 500, which is expected to bring increased bandwidth and additional compute dies. Both of these next-generation accelerators are slated for release in 2027.