According to Meta's official website, MTIA is an ASIC chip that can combine different circuits on the same chip and perform one or more tasks in parallel. The first generation MTIA ASIC chip was designed in 2020 and is an AI inference accelerator chip. The chip is manufactured using TSMC's 7nm process and operates at a frequency of 800 MHz. It has a computing power of 102.4 TOPS at INT8 accuracy and 51.2 TFLOPS at FP16 accuracy, with a thermal design power (TDP) of 25W.

 

In a blog article, Santosh Janardhan, Vice President and Infrastructure Manager of Meta, wrote that MTIA is Meta's internal customized accelerator chip for inferential work, which provides higher computing power and efficiency than CPUs and can provide customized services for internal work, achieving better performance, lower latency, and higher efficiency.

 

It is understood that over the past decade, Meta has spent billions of dollars hiring top data scientists to build new AI models and strive to lay out generative AI. Until 2022, Meta mainly used a combination of CPUs and chips designed specifically to accelerate AI algorithms to maintain the operation of its AI, and the task efficiency of this combination is usually lower than that of GPUs. This is also why Meta cancelled its original plan to promote CPU and AI custom chips on a large scale in 2022 and instead ordered billions of dollars worth of NVIDIA GPUs. In order to match these GPUs, Meta has undergone disruptive redesign of several of its data centers, which is also why Meta developed the MTIA ASIC chip, which is expected to be officially launched in 2025.