Meta, the social giant formerly known as Facebook, has been increasing its investment in artificial intelligence in recent years, not only in the metaverse, generative AI and other fields, but also independently developing hardware devices designed for AI to improve its computing power and efficiency and reduce its dependence on external vendors.
Recently, Meta's second-generation self-developed AI chip was officially put into production, this chip called Artemis, is a more advanced version of Meta's training and inference accelerator (MTIA), using advanced process technology, which can perform inference tasks in the data center, and work together with GPUs from Nvidia and other leading companies to provide strong support for Meta's various AI applications.
It is reported that Meta plans to deploy Artemis chips within this year to reduce its dependence on Nvidia GPUs. At present, Meta is already one of Nvidia's largest customers, and its CEO Zuckerberg has announced that he plans to deploy 350,000 Nvidia H100 GPUs by the end of this year, and a total of about 600,000 GPUs will run and train AI systems. A Meta spokesperson said: "We are confident that our self-developed accelerator will complement the GPUs on the market and provide the best balance of performance and efficiency for Meta's missions. ”
Meta's self-developed AI chips can not only improve the effectiveness of its recommender system, but also provide computing power for its own generative AI applications, as well as the GPT-4 open-source competitor LLAMA 3 that is being trained. Llama 3 is Meta's large language model that is planned to reach GPT-4 performance levels, but will still be available for free.
Another important advantage of Meta's self-developed AI chips is that they can be combined with its self-built AI supercomputer RSC to further improve its AI performance. RSC is a supercomputer assembled by Meta in cooperation with Penguin Computing, Nvidia and Pure Storage, and has completed the second phase of construction, including 2,000 NVIDIA DGX A100 systems and 16,000 NVIDIA A100 GPUs.
The launch of Meta's self-developed AI chip is undoubtedly a challenge to NVIDIA and an innovation in the field of AI. Meta's goal is very clear, which is to reduce its dependence on NVIDIA chips while controlling the cost of AI tasks as much as possible, paving the way for its vision of the metaverse and generative AI.
What do you think of Meta's self-developed AI chip? Welcome to leave a message in the comment area and join us**.