The launch of the latest version of MTIA marks a significant turning point by showing substantial improvements in performance, a key factor that will help power Meta's ad ranking and recommendation models. This advancement not only demonstrates the technical evolution of AI-specific hardware, but also underlines Meta's commitment to investing resources in the development of cutting-edge AI infrastructure. On its own, without necessarily resorting to third-party technological services or infrastructures, rather trying to develop a dominant strategy in the evolution of AI technologies. Meta designed the architecture of this chip with the intent to provide the right balance between compute capacity, memory bandwidth and memory capacity to serve ranking and recommendation models.
The accelerator comes with an 8×8 grid of processing elements (PEs). These PEs can produce significantly increased dense computing performance (3.5x compared to MTIA v1) and sparse computing performance (7x improvement). To achieve this performance, Meta tripled the size of the PE's local memory, doubled the on-chip SRAM and increased its bandwidth by 3.5 times, and doubled the capacity of LPDDR5. The race for AI leadership has just begun.
#Meta39s #nextgeneration #training #inference #accelerator
