Yesterday, in a fascinating display of technological innovation, Meta revealed plans to increase its emphasis on artificial intelligence (AI) and the development of custom processors. During a virtual event, Alexis Bjorlin, VP of Infrastructure at Meta, emphasized the company’s vertical integration approach to advancing AI research at scale.
Despite being a relative latecomer in adopting AI-friendly hardware systems, Meta has taken deliberate steps to enhance its AI capabilities. Over the past decade, the company has invested billions in AI talent and technology, utilizing AI for various services, including content moderation, discovery engines, and advertising recommenders.
Nevertheless, the past was marred by setbacks. Meta’s custom chip for accelerating AI algorithms, slated for a 2022 launch, was shelved, prompting the company to pivot towards Nvidia GPUs. Several data centers required extensive modifications due to this transition.
Meta now intends to regain control by developing its chip, Meta Training and Inference Accelerator (MTIA), with a 2025 release date. The MTIA is a component of Meta’s effort to create a suite of processors optimized for AI training and inference workloads. The MTIA, based on an ASIC design, enables the concurrent execution of multiple duties, thereby enhancing efficiency.
Bjorlin emphasized the importance of MTIA in obtaining greater efficiency and performance across their most important workloads. This action aligns Meta with other tech titans such as Google, Amazon, and Microsoft, who have also employed custom AI chips to enhance their AI capabilities.
The MTIA is believed to significantly improve Meta’s performance per watt when executing recommendation workloads, even though memory and networking still have space for improvement.
In parallel, the Research SuperCluster (RSC), a supercomputer dedicated to research, drives Meta’s AI research initiatives. Built-in collaboration with Penguin Computing, Nvidia, and Pure Storage, the RSC is Meta’s response to its competitors’ supercomputing efforts.
The RSC is a training platform for researchers to create models based on examples from Meta’s production systems. It is among the world’s most powerful supercomputers, with nearly five exaflops of computing capacity. The RSC was instrumental in training “Large Language Model Meta AI” (LLaMA), one of Meta’s large language models.
Meta Scalable Video Processor (MSVP), a chip designed to handle video-on-demand and live broadcasting, was also revealed in the announcement made today. This change is in response to an increase in Facebook video consumption, which accounts for half of the app’s total utilization time.
Overall, these hardware advancements emphasize Meta’s urgent efforts to expedite the development of generative AI capabilities. With Goldman Sachs predicting a $150 billion market for generative AI software, even a fraction of this market could compensate for Meta’s losses in other areas, such as augmented reality headsets and VR platforms.
Meta’s strategic shift towards AI and custom processor development is an aggressive move to maintain its competitive position in the digital landscape. Time will tell whether these initiatives can produce the transformative results that Meta envisions.