Meta's new MTIA 400 chip has a split personality: Training AI and serving ads
Faster than Blackwell, but still no replacement for AMD or Nvidia ... yet
Meta's latest AI accelerator, the MTIA 400, boasts a dual-purpose design. While it can handle AI inference tasks, its primary function is to power Meta's ad recommender systems. The chip, a product of Meta's MTIA (Meta Training and Inference Accelerator) series, is built on a heterogeneous multi-die architecture. This means it is constructed from multiple chips, each specialized for a specific task.
The MTIA 400 features two compute chiplets, two I/O chiplets, and an SoC die, all working together to deliver impressive performance. With a 6x8 grid of processing elements and a 3nm process technology, the chip can deliver 12 petaFLOPS of compute power at 1.7 GHz. Memory-wise, it is equipped with eight 36 GB HBM3e stacks, providing 288 GB of memory and 9.2 TB/s of bandwidth.
Despite its impressive capabilities, the MTIA 400 is still 3x to 3.3x slower than Nvidia's top-specced Blackwell accelerators and between 3x and 3.3x slower than AMD's Instinct MI455X. However, its superior memory bandwidth and faster performance make it a strong contender for training large language models (LLMs). Meta is also working on an inference-optimized version of the chip, the MTIA 450, which will have double the memory bandwidth and is expected to enter production next year.
Written by urgent.news from The Register Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
