At the company's developer conference, Nvidia introduces a new Tesla GPU and computing module aimed at driving innovation in artificial intelligence.
SAN JOSE, Calif.—Nvidia executives five years ago decided to put a lot of the GPU maker's resources into developing technologies for the artificial intelligence market.
There were developments in the space—including the development of new algorithms that made it possible to use deep learning techniques to drive innovation around artificial intelligence (AI)—though there were no guarantees that the market would grow, according to Jen-Hsun Huang, co-founder and CEO of Nvidia.
However, efforts by a growing number of major tech players—including Google, Microsoft and Baidu—have driven innovation in AI, making 2015 the most significant year for the space, Huang said April 5 during the opening keynote address at Nvidia's GPU Technology Conference (GTC) 2016 here. Much of that development was done using GPUs. During his talk, Huang unveiled some of the results of what Nvidia has been working on during those years.
During his talk, Huang also announced new products in the areas of virtual reality and connected cars, which is combining with deep learning to be the three pillars of the GTC event this week.
The CEO showed off the Tesla P100, a massive chip based on Nvidia's new 16-nanometer Pascal architecture that—when counting the GPU cores and memory on the chip—packs 150 billion transistors that was built for data center and cloud environments. Nvidia's history has been building GPUs for PCs and gaming systems, but the company is pushing to expand the use of GPUs in enterprise and cloud computing, beyond the GPU accelerators found in many high-performance computing systems.
The highly dense chip, aimed at hyperscale environments, also uses the 16nm FinFET manufacturing process to help drive performance while keeping down power consumption, 16GB of second-genration High-Bandwidth Memory (HBM2) technology—in which the dies are stacked—and Nvidia's NVLink interconnect technology to connect with other Tesla P100 modules.
In addition, Huang introduced the DGX-1, calling it the world's first supercomputer for deep learning and AI that combines eight of the Tesla P100 GPUs with two Intel Xeon server chips to drive 170 teraflops of performance in a 3U (5.25-inch) form factor. It offers 12 times the performance of the system Nvidia introduced at last year's GTC that was based on the 28nm Maxwell architecture. It will come priced at $129,000 and carry a 3,500-watt power budget, but Huang also said it will offer the same performance as 250 CPU-based servers.
"Our company has gone all-on in deep learning," he said, adding that AI represents "a brand new computing model. … Our strategy is to accelerate deep learning everywhere."
Deep learning uses layers of nodes that process data coming into the system, with the idea that as the data passes through each node, the system gets closer to the correct result to whatever question is being asked. These neural networks also are designed to enable artificial intelligence so the system can learn from its experience, much like a human brain does. AI is being used now for such work as image recognition, but Huang and other Nvidia executives said the use is almost limitless.
By offering a fully integrated platform like the DGX-1, the company has made deep learning and AI technologies available to a broad range of customers, the CEO said.
"Deep learning has been democratized," he said.
Huang said the Tesla P100 chips are in volume production now and will begin shipping "soon." He expects the major cloud service providers—such as Facebook, Google and Microsoft—to be the first to leverage the massive GPUs in their environments, with OEMs getting the chips later in the year and rolling out the first systems using the P100 in the first quarter 2017. Dell, Hewlett Packard Enterprise and Lenovo already have been using samples of the P100 to begin developing systems, he said.
The DGX-1 will first be sent to the top researchers in AI, including Stanford University, Massachusetts Institute of Technology (MIT), New York University and the University of Toronto.
Nathan Brookwood, principal analyst with Insight 64, told eWEEK
he was impressed with what Huang showed off during the keynote address, not only in regards to the technology but also the fact that the products are real and about to get into the market. It's good for both the industry and for Nvidia, he said.
"They've clearly made a major thrust [behind AI] in the company," Brookwood said. "When they first talked about it a few years ago, it wasn't clear how things [in the market] would play out. They made some very heavy bets and it now appears that they will be able to collect."