NVIDIA’s GPU Technology Conference (GTC) opened this morning with an interesting concept. What would happen if you could create a machine that could spark innovation? This is the biggest push toward artificial intelligence (AI) since the launch of IBM’s Watson. NVIDIA is announcing five big things at this show, most of which have to do with scaling intelligent machines down so that everyone could gain access to them. Note that this conference has doubled since 2012 and it is now held worldwide.
At this show, NVIDIA is announcing a new processor, a new server, and an updated automotive platform, all designed to drive deep learning and machine intelligence, across the market.
Let’s talk about some of the highlights.
AI Processor-Tesla P100
Leading into this announcement, NVIDIA refreshed on a joint project between Google and Microsoft, ImageNet, which sees images better than humans; Microsoft’s “Super Deep Network,” which was 152 layers deep; the Berkeley Brett, a robot that could easily be trained to do a wide variety of simple tasks; Baidu’s Deep Speech, which does translation in real time; and AlphaGo, the first demonstration of a computer that can beat a human Go expert. (It should be noted that it was estimated that it would take another 100 years from the time when a computer beat a chess master before a computer could do the same with a Go champion.) This all showcases that we are at a point where computers can perform tasks that we can’t write software for; it is the age of deep learning.
NVIDIA spoke to Facebook’s AI research and demonstrated how it could rapidly train a computer using 20,000 images. After the training, the computer was able to create pictures that looked painted and conformed to the paintings it had analyzed. For instance, given the command to generate a sunset on the beach, it created one that didn’t appear abstracted from the pictures it had seen but was undeniably a good painting of a beach during a sunset. This action used “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.”
This built up to the launch of the Tesla P100, a 150B transistor part specifically designed for AI and hyperscale computing. This can perform at 21.2 teraflops. To create this part, NVIDIA needed to solve five problems, each of which was thought to be incredibly hard.
NVIDIA DGX-1 Deep Learning Computer
NVIDIA then moved on to announce its deep learning server, the DGX-1, a 170 Teraflop monster that was designed from the ground up to do deep learning. On comparison, using the existing technology, your fastest is now 3 TF, training time for Alexnet is 150 hours, and if you wanted to cut the time to two hours you’d need 250 servers. The DGX-1 will do Alexnet in two hours by itself. This reflects a 12X performance improvement year over year. This is AI in a box. For a sense of what this means, this computer costs $129,000; just to network the 250 servers that this replaces would cost $500,000.
The first receivers of this technology will be the research organizations leading the research efforts in AI, particularly in health care research.
Deep Learning for Cars
NVIDIA has reached a major expansion of its efforts for self-driving cars. The NVIDIA Drive PX is the first commercial AI computer for cars. At the heart of this presentation was the program DaveNet, which teaches cars how to drive the way you teach people to drive, by driving. NVIDIA announced on top of this Roborace, an autonomous driving race using specially built electric cars that drive themselves.
Really the big part of this is the realization that it will be both faster and better to teach cars how to learn so they can teach themselves and to adapt several layers of problem solving. So the common script-based method of see, plan, react would be supplemented and enhanced by information that the increasingly deployed self-driving cars would be learning on the road as well as specific vehicle training. In short, not only will the cars start out smarter, they will get even smarter at an inhuman rate of speed.
Wrapping Up: Robots Now Tread Where Only Humans Once Did
From computers that can create art to those that can drive race cars, we are seeing what is clearly a massive surge of computers with the capability to make real-time decisions without human interaction. From recognizing threats that drivers don’t see to being able to create original art by understanding what constitutes “art,” these machines will increasingly move in places where only humans did before. Clearly, social, societal and legal changes will need to be made as these systems grow ever faster and more capable. Suddenly, I’m feeling a tad obsolete.
In the end, here at GTC, NVIDIA has made a major jump from using technology to create entertainment to using it to alter our reality and change the world. I think it likely we’ll look back at 2016 and recognize it as the last year it was clear that humans were at the top of the food chain and where there were limits to what robots could do.
This is the beginning of a whole new world.
Rob Enderle is President and Principal Analyst of the Enderle Group, a forward-looking emerging technology advisory firm. With over 30 years’ experience in emerging technologies, he has provided regional and global companies with guidance in how to better target customer needs; create new business opportunities; anticipate technology changes; select vendors and products; and present their products in the best possible light. Rob covers the technology industry broadly. Before founding the Enderle Group, Rob was the Senior Research Fellow for Forrester Research and the Giga Information Group, and held senior positions at IBM and ROLM. Follow Rob on Twitter @enderle, on Facebook and on Google+