According to official Meta information, Meta is building the next generation of large-scale infrastructure with AI at its core, which includes supporting new generative AI product services, recommender systems, and conducting advanced AI research. We expect this investment to continue to expand in the coming years as AI models become increasingly complex and demand for computing power increases.
Last year, Meta launched its first generation of AI inference gas pedals, the Meta Training and Inference Accelerator (MTIA) v1. This gas pedal was developed in-house by Meta and is designed specifically for the Company's AI workloads, specifically for deep learning recommendation models, which are improving the diversity of experiences in our products .
The MTIA program is a long-term investment by Meta to provide the most efficient architectural solution to meet the needs of its unique workloads. As AI workloads become more and more critical in our products and services, this efficient performance will enhance the ability to deliver superior experiences to our users around the world.MTIA v1 is an even more significant step forward in improving the computational efficiency of our infrastructure and better supporting software developers in building AI models that enable new and better user experiences.
This follows Meta's promise that the next generation of custom AI chips will be more powerful and able to train its ranking models faster. The company said in a post that MTIA is a big piece of its long-term plan to build infrastructure around how AI is used in its services. Its chief said, "Realizing our ambitions for custom silicon means investing not only in compute silicon, but also in in-memory broadband, networking and capacity, and other next-generation hardware systems."
Meta announced MTIA v1 in May 2023 with a focus on delivering these chips to data centers. The next generation of MTIA chips may similarly target data centers.MTIA v1 wasn't originally expected to be released until 2025, but Meta says two MTIA chips are now in production.
Currently, MTIA is primarily used to train ranking and recommendation algorithms, but Meta said its goal is to eventually expand the chip's capabilities to begin training generative AI, such as its Llama language model.
Meta emphasized that the new MTIA chip is "essentially focused on achieving the right balance between compute power, memory bandwidth and memory capacity." This new chip will come with 256MB of on-chip memory and a 1.3GHz clock frequency compared to the first generation's 128MB of memory and 800GHz frequency.
As the demand for computing power grows with the increase in AI applications, other companies in the AI space are exploring making their own chips. Google unveiled its new TPU chip in 2017, Microsoft announced its Maia 100 chip, and Amazon introduced its Trainium 2 chip, which trains four times faster than its predecessor.
The scramble for high-performance chips underscores the importance of custom chips in running AI models. With the surge in demand for chips, NVIDIA, which currently dominates the AI chip market, has a market capitalization of $2 trillion. Meta has been working on the MTIA v2 project for quite some time. It has previously been reported that it is focusing on inference capabilities, aiming to improve the performance and efficiency of AI applications. The MTIA chips are built using an advanced manufacturing process, the latest of which utilizes a 5nm process, which is a step up from its predecessor's 7nm process. In addition, Meta says that the new MTIA chip delivers three times the performance of its predecessor, demonstrating significant progress in the AI space and a commitment to improving its own technological prowess. With this chip, Meta hopes to reduce its dependence on other chip suppliers such as NVIDIA and strengthen its own competitiveness in the AI field. The new MTIA chip has already been deployed in Meta's data centers and has begun serving AI applications.
