Meta announced the next-generation Meta Training and Inference Accelerator chip (MTIA), designed for Meta’s AI workloads.
Meta released the first-generation MTIA chips last year. The company’s new-generation version Offers significantly enhanced performance.
This next-generation MTIA product is part of a broader, full-stack program to develop custom silicon for our specific workloads and system requirements. The new MTIA version more than doubles our existing solution’s compute and memory bandwidth, while still maintaining our tight integration with our workloads. The system is built to provide efficient ranking and recommendations models to give users high-quality suggestions.
The architecture of this chip is based on a fundamental focus to provide the best balance between compute power, memory bandwidth and storage capacity in order to serve ranking and recommendation algorithms.
Meta claims that the MTIA chip excels in both high and low complexity models of ranking and recommendations. This is a critical element to Meta’s company. According to the company, controlling its entire stack is an advantage compared with using standard GPUs.
Our custom silicon is designed to be compatible with both our current infrastructure and with future hardware, such as next-generation GPUs. To meet our ambitious goals for custom silicon, we must invest not only in computing silicon, but also memory bandwidth, network capacity and other future-generation hardware.