Meta has launched an upgraded version of its Meta Training and Inference Accelerator (MTIA), achieving substantial enhancements in compute and memory capabilities. The next-generation MTIA offers a threefold increase in dense compute performance and a sevenfold improvement in sparse compute operations, along with significantly expanded memory capacity and bandwidth. These advancements not only boost the accelerator’s efficiency and capability in handling complex AI tasks but also mark a strategic move towards reducing reliance on external hardware suppliers like NVIDIA.