Facebook Open Sources ‘Big Sur’ AI Hardware Design.
Facebook is releasing the hardware design for a server it uses to train artificial intelligence software, allowing other companies exploring AI to build similar systems. Facebook Inc.’s use of artificial intelligence, which ranges from tools for image recognition to the filtering of the news feeds for its social network, demands special computing infrastructure.
NVIDIA (NASDAQ: NVDA) today announced that Facebook will power its next-generation computing system with the NVIDIA® Tesla® Accelerated Computing Platform, enabling it to drive a broad range of machine learning applications.Over the last few years, a technology called deep learning has proven so adept at identifying images, recognizing spoken words, and translating from one language to another, the titans of Silicon Valley are eager to push the state of the art even further—and push it quickly.
Facebook said today it has developed a new computing system aimed at artificial intelligence research that is twice as fast and twice as efficient as anything available before.Facebook wants to speed up research into artificial intelligence for everyone by making the plans for a massively powered box available to any company that wants to hasten up its efforts to build better facial or voice recognition. Code-named Big Sur, Facebook uses the server to run its machine learning programs, a type of AI software that “learns” and gets better at tasks over time. The company recently began building custom servers for its artificial intelligence workload and Thursday announced it would release the designs for that powerful hardware to the world — for free.
While training complex deep neural networks to conduct machine learning can take days or weeks on even the fastest computers, the Tesla platform can slash this by 10-20x. These days, machine learning and artificial intelligence are, hand in hand, becoming the lifeblood of broad new applications throughout the business and research communities. 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. The company said the plan to open-source the blueprints of the servers — called “Big Sur” — would help other companies and researchers benefit from the incessant tweaking of Facebook’s developers. At Google, this tech not only helps the company recognize the commands you bark into your Android phone and instantly translate foreign street signs when you turn your phone their way.
But even as that dynamic has been significantly driven by computers that are more powerful and more efficient, industry is reaching the limits of what those computers can do. Increasingly, Facebook is developing elements of its business centered on artificial intelligence, and the social networking giant’s ability to build and train advanced AI models has been tied to the power of the hardware it uses. At Facebook, it helps identify faces in photos, choose content for your News Feed, and even deliver flowers ordered through M, the company’s experimental personal assistant. Among its recent AI projects have been efforts to make Facebook easier to use for the blind, and to incorporate artificial intelligence into everyday users’ tasks. And because it’s contributing these designs to the Open Compute Foundation, much like it has done for its server designs and networking gear, other companies can take these designs and build their own AI hardware or even tweak the Big Sur designs to make them better.
All the while, these two titans hope to refine deep learning so that it can carry on real conversations—and perhaps even exhibit something close to common sense. GPUs are widely used in artificial intelligence because the chips have far more individual processing cores on them than traditional processors produced by Intel Corp., making them adept at the dumb-but-numerous calculations required by AI software. 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. Our goal is to provide researchers and companies with the most productive platform to advance this exciting work.” In addition to reducing neural network training time, GPUs offer a number of other advantages. Serkan Piantino, director of engineering at Facebook’s AI Research, said on a conference call explaining the news, that because of the enormous amount of heat and power drawn by the graphics processing chips used in building the machines, that the team that melted its first enclosure would have gotten a steak dinner.
That’s because Piantino told reporters, “our capabilities keep growing, and with each new capability, whether it’s computer vision, or speech, our models get more expensive to run, incrementally, each time.” Also, he said, as the FAIR group has moved from research to capability, it has seen product groups from across Facebook reach out about collaborations. And the Tesla platform’s growing global adoption facilitates open collaboration with researchers around the world, fueling new waves of discovery and innovation in the machine learning field. Yann LeCun, the head of AI research at Facebook, said that negotiating with the data center administrators to take in the necessary power to ensure the machines got the juice they needed was a significant issue.
Big Sur Optimized for Machine Learning NVIDIA worked with Facebook engineers on the design of Big Sur, optimizing it to deliver maximum performance for machine learning workloads, including the training of large neural networks across multiple Tesla GPUs. For Facebook, releasing its designs has potent benefits: the openness can be a major incentive for top talent to join the company; firms that use the equipment may contribute their improvements back to the community, letting Facebook outsource some of its research and development costs; and if enough people buy the equipment, then economies of scale will ultimately lower the price Facebook pays for its computer hardware, Serkan Piantino, the engineering director of Facebook’s AI group, said in an briefing with reporters. “Often the things we open-source become standards in the community and it makes it easier and cheaper for us to acquire the things later because we put them out there,” Piantino said.
It hasn’t released the hardware specs yet, but images show a large airflow unit inside the server that presumably contains fans that blow cool air across the components. Keep Current on NVIDIA Subscribe to the NVIDIA blog, follow us on Facebook, Google+, Twitter, LinkedIn and Instagram, and view NVIDIA videos on YouTube and images on Flickr. OCP members are fond of talking about the “gratuitous differentiation” that server vendors put in their products, which can drive up costs and make it harder to manage equipment from different vendors. “We’ve removed the components that don’t get used very much, and components that fail relatively frequently — such as hard drives and DIMMs — can now be removed and replaced in a few seconds,” Facebook said. The company’s technologies are transforming a world of displays into a world of interactive discovery — for everyone from gamers to scientists, and consumers to enterprise customers. All the handles and levers that technicians are supposed to touch are colored green, so the machines can be serviced quickly, and even the motherboard can be removed within a minute. “In fact, Big Sur is almost entirely tool-less –the CPU heat sinks are the only things you need a screwdriver for” Facebook says.
The Internet’s largest services typically run on open source software. “Open source is the currency of developers now,” says Sean Stephens, the CEO of a software company called Perfect. “It’s how they share their thoughts and ideas. In the closed source world, developers don’t have a lot of room to move.” And as these services shift to a new breed of streamlined hardware better suited to running enormous operations, many companies are sharing their hardware designs as well. Important factors that could cause actual results to differ materially include: global economic conditions; our reliance on third parties to manufacture, assemble, package and test our products; the impact of technological development and competition; development of new products and technologies or enhancements to our existing product and technologies; market acceptance of our products or our partners’ products; design, manufacturing or software defects; changes in consumer preferences or demands; changes in industry standards and interfaces; unexpected loss of performance of our products or technologies when integrated into systems; as well as other factors detailed from time to time in the reports NVIDIA files with the Securities and Exchange Commission, or SEC, including its Form 10-Q for the fiscal period ended October 25, 2015. These forward-looking statements are not guarantees of future performance and speak only as of the date hereof, and, except as required by law, NVIDIA disclaims any obligation to update these forward-looking statements to reflect future events or circumstances. © 2015 NVIDIA Corporation.
Although GPUs were originally designed to render images for computer games and other highly graphical applications, they’ve proven remarkably adept at deep learning. Traditional processors help drive these machines, but big companies like Facebook and Google and Baidu have found that their neural networks are far more efficient if they shift much of the computation onto GPUs. In short, Facebook can achieve a greater level of AI at a quicker pace. “The bigger you make the neural nets, the better they will work,” LeCun says. “The more data you get them, the better they will work.” And since deep neural nets serve such a wide variety of applications—from face recognition to natural language understanding—this single system design can significantly advance the progress of Facebook as a whole.