The most surreal and deeply fascinating moment of Computex 2017 took place yesterday during the Nvidia press conference, when over a dozen virtual robots tried to slap hockey pucks into nets. It was a demonstration of how autonomous cars will learn how to drive themselves in the future. We can explain!
Nvidia CEO Jensen Huang was using this simple demo to outline the Isaac Initiative. Named for Isaac Asimov, Nvidia’s ambitious AI program combines all the elements of their suddenly breakneck advances in artificial intelligence — robot hardware, machine learning software, new ultra-powerful GPUs, and a fascinating virtual reality experiment they announced this week called Isaac’s Lab.
It’s elegant how all these factors come together — elegant enough to tell you that Nvidia has probably been planning this for a very, very long time. The key is their new Volta GPUs, which were announced earlier this month (starting with the Tesla V100 below). Don’t expect these to be gaming optimized GPUs anytime soon, either — these things are on a whole other level. The massive GPUs have 21 billion transistors, 640 tensor cores, and can put out over 100 Teraflops per second. That blows past Nvidia’s Pascal architecture from last year, as well as AMD’s Vega architecture planned for this year. Little wonder — Huang said Nvidia invested $3 billion in developing Volta.
What’s the point of all that power? Well, virtual reality is kind of underwhelming right now. Because it demands so much processing power, the environments end up being low-resolution and lacking in detail. That, and physics can be all over the place, depending on the game. Volta has been designed to deliver photorealism and real-world laws of physics in a virtual reality environment, which explains the gaudy numbers above.
If you’re a gamer and you’re salivating over that prospect, you might want to pump the brakes a bit. While this might change in the future, right now, Volta isn’t meant for gamers. Honestly, it’s not even meant for people — Nvidia is doing it for the machines.
Here’s where everything comes together. Currently, any AI we interact with (like Google Assistant), is disembodied. The next step is to take those AIs and stick them in functioning robots that can do tasks by themselves, giving you a Siri bot that could someday, I don’t know, make you some pancakes. Trouble is, you need to train AI to do those tasks, and training an AI while it’s in an active robot could be dangerous.
Autonomous cars are the perfect example. If a car is driving itself, it’s essentially an AI-powered robot. And, like with any human, you wouldn’t want to throw that AI on the road and have it learn on the fly. So, Nvidia is using the power of Volta to create virtual representations of the world that obey the laws of physics, basically creating a driving simulator for automotive AIs. AI developers can then fill up that virtual world with simulated driving situations until those AIs learn how to drive. Once training is done, the AI can be taken from the cloud to a real car ready to drive on real roads.
Huang even talked about speeding up the learning process by having multiple AIs learn, then watching for the best students. The top AI would then be copied, replacing all the other slower AIs, and the training program would then resume. Was there nervous laughter from the crowd? Oh, you bet there was.
That brings us back to the hockey example. It was a simple demo of how the learning process works — the same way it works for humans. Those hockey-playing AIs received positive signals when they scored, and negative signals when they missed. Over time, the AIs will analyze what works and what doesn’t, finally becoming perfect goal scoring machines.
If you replace ‘goal scoring’ in that sentence with some other things, then it starts to sound a little Skynetty. Nvidia actually started getting Skynet jokes when they built a supercomputer out of their P100 Pascal GPUs, and considering that similarly configured Volta GPUs blow the doors off those in terms of performance, it’s only a matter of time before parents start naming their kids John Connor on purpose.
Check out more of our Computex 2017 coverage right here!
Disclosure: Taiwan External Trade Development Council (TAITRA), an organizer of COMPUTEX, covered our travel expenses to attend COMPUTEX 2017. All thoughts and opinions are 100% our own.