Meta announced that it will rebuild its next-generation infrastructure architecture, covering both hardware and software stacks, to enhance the development of artificial intelligence technology and enable more efficient deployment of new technologies, thereby driving the development of future metaverse applications through artificial intelligence.
The next-generation infrastructure designed for AI will include Meta's first customized chip for running AI models, a new data center design optimized for AI operations, and a supercomputer equipped with 1 GPUs for accelerated computing.
At the same time, Meta emphasizes that artificial intelligence is the core of its products. It not only enhances the level of personalized experience, develops safer and fairer products, and creates a richer experience, but also helps businesses reach the audiences they value most.
Next, Meta plans to adjust the way programming is done by using Code Compose, an internally developed generative artificial intelligence programming assistance tool, to improve developers' work efficiency throughout the software development life cycle.
Since establishing its first data center in 2010, Meta has continued to advance its infrastructure through three key initiatives: Big Sur hardware in 2015, the development of the PyTorch programming language, and last year's supercomputer designed for artificial intelligence research.
• MTIA (Meta Training and Inference Accelerator): MTIA is Meta's first family of custom accelerator chips developed in-house, specifically for inference-related workloads. Designed for in-house workloads, MTIA delivers superior computing performance and processing efficiency compared to CPUs. By deploying MTIA chips alongside GPUs, performance for each workload is improved, latency is reduced, and processing efficiency is boosted.
• Next-generation data center:Meta's next-generation data center design not only supports existing products but will also facilitate training and inference for future AI hardware. Designed for AI optimization, this new data center supports liquid-cooled AI hardware and a highly efficient AI network, connecting thousands of AI chips to form data center-scale AI training clusters.
Both development time and cost will be increased, and it can also complement other new hardware devices, such as Meta's first internally developed ASIC solution, the MSVP (Meta Scalable Video Processor), to support the continued growth of audio and video content.
• Research SuperCluster (RSC) artificial intelligence supercomputer:Meta's RSC is one of the world's fastest AI supercomputers, capable of training the next generation of large-scale AI models to support new augmented reality (AR) tools, content understanding systems, real-time translation technologies, and more. It is equipped with 1 GPUs, all accessible through a three-layer Clos network structure, providing sufficient resources for each of the 6000 training systems.
Since last year, RSC has been involved in various research projects, such as the Large Language Model Meta AI (LLaMA) promoted by Meta and announced earlier this year.
Furthermore, Meta's custom-designed infrastructure will enhance the end-to-end user experience across the physical, virtual, and software layers, as well as the user experience. From the data center to the server equipment, and even the mechanical systems that maintain all operations, Meta designs, develops, and operates entirely in-house. This allows Meta to control the entire architecture from top to bottom and design it based on Meta's specific needs.
Meta anticipates that large-scale, self-built infrastructure will become increasingly important in the future. In the next 10 years, we anticipate seeing more customized chip designs, domain-specific AI architectures, and new systems and tools designed for large-scale deployments. These will enable Meta to develop more complex AI models based on the latest research findings.
