Over the past several years, Facebook has developed software that can read stories, answer questions, play games, and even learn things. Now, the company is taking its artificial intelligence efforts to the next level: hardware.
The social network on Thursday unveiled new AI hardware its researchers have developed to train neural networks, and open-sourced the design, offering other organizations a blueprint for how to set up their own AI-specific infrastructure. Designed in collaboration with Nvidia, the Open Rack-compatible hardware is designed for AI computing at a large scale.
Dubbed “Big Sur,” the hardware includes eight high-performance GPU boards of up to 33 watts each. Compared to Facebook’s previous system, “we can train twice as fast and explore networks twice as large,” Facebook’s Kevin Lee and Serkan Piantino wrote in a blog post.
“Distributing training across eight GPUs allows us to scale the size and speed of our networks by another factor of two,” they added.
It’s also “far more versatile and efficient” than the off-the-shelf solutions Facebook was using in the past. Many high-performance computing systems require special cooling to operate, but Facebook has designed its new servers for “thermal power and efficiency,” allowing the company to operate them in its own free-air cooled, Open Compute standard data centers. Facebook also ditched components that don’t get used very much and made it so that components that fail often — like hard drives — can be removed and replaced within “seconds.”
“We want to make it a lot easier for AI researchers to share techniques and technologies,” Lee and Piantino wrote. “As with all hardware systems that are released into the open, it’s our hope that others will be able to work with us to improve it.”
Meanwhile, Facebook also on Thursday said its AI Research team is more than tripling its investment in GPU hardware, and enabling other teams across the company to use neural network in their products and services.