Greetings, crazy night owls/early birds! In a little under an hour, Nvidia will officially kick off the GPU Technology Conference in San Jose, California with its opening keynote address. Confirmed speakers include Nvidia CEO Jen-Hsun Huang (natch) and Tesla Motors CEO Elon Musk. Keep updating this page for all the biggest announcments!
All times are in AEDT.
2:30am, 18 March 2015
Guests are slowly filtering into the San Jose McEnery Convention Center. Only 30 minutes to go! So what do you guys think will be announced? In an unusual move, Nvidia chose to unveil its eight billion transistor-strong GeForce GTX Titan X card at GDC 2015 instead of its own GPU conference. Presumably, this means it has a few surprises up its sleeve.
2:45am, 18 March 2015
It’s currently Saint Patrick’s Day in America (they’re a day behind us dontchaknow): who wants to bet that this will be mentioned onstage in reference to Nvidia’s green logo?
3:05am 18 March 2015
Er, any moment now…
3:15am 18 March 2015
And we’re off! First up is a video celebrating cinematic visions of the future including Blade Runner, Star Trek, Flash Gordon and Forbidden Planet interspersed with emerging tech from Nvidia’s own labs. Looks like robots are going to feature prominently…
3:19am 18 March 2015
Dispensing with an MC, Jen-Hsun Huang takes straight to the stage in his favourite new leather jacket. On today’s agenda: a new GPU, deep learning, Nvidia’s roadmap, deep learning, self-driving cars and deep learning. “We have a lot to tell you about deep learning,” Huang joked.
3:20am 18 March 2015
According to Huang, cars with dials, knobs and levers are now a thing of the past. We’d argue over the knob one — plenty of those still driving around.
3:22am 18 March 2015
Now he’s crowing about parallel computing platform CUDA. 3 million downloads. 800 universities currently teaching the programming model worldwide.
3:26am 18 March 2015
Say hello to the Titan X! Based on Nvidia’s 28nm GM200-400 Maxwell GPU (AKA the fastest chip Nvidia has ever produced), its sheer processing grunt is equivalent to two flagship GK110 cores from the previous generation. Yikes. It features six graphics processing clusters (GPCs) that each boast four Maxwell Streaming Multiprocessor (SMMs) for a total of 24. It also comes with 192 Texture Mapping Units (TMUs) and 96 Raster Operation Units (ROPs). The GPU, meanwhile, packs in six 64-bit memory controllers, for a 384-bit memory interface.
To give an idea of what the Titan X is capable of, we were shown an animated movie created by Epic Games with Unreal Engine 4. It featured a kid chasing a kite across a mountain landscape and through photo-realistic water. “This story simply wasn’t possible to tell without the advancement of computer graphics.”
3:40am 18 March 2015
Titan X isn’t just about graphics. During testing, it knocked deep structured learning in machines down from 43 days to just three. In other words, it’s a high-end gaming card that science boffins can get use out of too.
3:43am 18 March 2015
We have a price! $US999. Thanks to the speed gains, Huang reckons the card will pay for itself in a day. (If you happen to be a deep learning network researcher, that is.)
3:45am 18 March 2015
And we’re back to deep learning. Basically, a history lesson. It’s being likened to the “Big Bang” of computer perception.
3:55am 18 March 2015
It wouldn’t be a technology conference without a technical hitch. Visuals are currently down.
3:58am 18 March 2015
We’re now getting deep into the visual recognition prowess of Imagenet. The 19 layer-deep network can train itself to recognise a subspecies of animal that is presented to it by filtering through millions of photos. Thanks to deep learning, image search is about to get a whole lot more sophisticated.
4:08am 18 March 2015
According to Huang, deep learning models are currently revolutionising medical research. “Predicting cancer is about predicting the growth of cancer cells. Deep learning can analyse the patterns by itself and do an even better job [than doctors].” Recent success stories include detecting mitosis in breast cancer cells, assessing the toxicity of new drugs and understanding gene mutations.
4:22am 18 March 2015
Here’s a new product aimed exclusively at developers and researchers: the Digits Devbox. It boasts the world’s fastest GPU, multi-GPU training and interface and the maximum GPU processing power out of a plug. It will be available May 2015 for $15,000. By Huang’s own admission, Nvidia isn’t expecting to sell a lot of these.
4:27am 18 March 2015
We’re now onto Nvidia’s next-gen Pascal GPU which boats mixed precision, more memory bandwidth and NVLink. It’s 28 times faster than the fastest supercomputer in the year 2000 and 10 times faster than Nvidia’s Maxwell chip. Hnngh.
4:35am 18 March 2015
“Driving isn’t about detecting, it’s about learned behavior.” We’re now onto Nvidia’s Drive PX self-driving car computer. Deep learning architecture will allow cars to assess and react to every possible driving condition far more efficiently than a human driver.
“How do you teach a baby to play ping pong? Teaching it physics is one way…Or you could just show them by hitting a couple of balls in front of them and then putting a bat in their hands. Every time they hit the ball correctly, you reinforce the behavior. Teaching a baby to play ping pong and teaching a car to drive itself are quite similar.”
4:43am 18 March 2015
We’re now watching a video of DAVE (DARPA Automative Vehicle) in action — it’s successfully traversing an obstacle-filled environment with all decisions powered by a single CPU. The car was trained purely through training data consisting of 225k images. Its behavior and the sophistication of its driving abilities gradually improved with the more images it was given access to. By the end, it was capable of making the same decisions a human driver would.
As impressive as DAVE was, it’s nothing compared to Nivida’s Drive PX self-driving car platform: it has 630 connections that can be fired off at 185 frames per second. Developers can pick up the Drive PX kit for a cool $10,000. It will be available from May, with the first prototype cars being deployed shortly thereafter.
4:53am 18 March 2015
It’s Tesla CEO Elon Musk! Huang quizzed Musk about his noted trepidation of A.I computing, which he once claimed could be more dangerous than nuclear warfare. Apparently, autonomous cars are okay as it’s a focused platform working towards a basic, singular purpose. (Evidently, he hasn’t seen the killer-car movie Christine.)
As you would expect, speed is the current stumbling block when it comes to self-driving cars. The ceiling for safe driving is currently 5-10 miles per hour, although highway cruising at accelerated speeds is also easy as there are fewer environmental factors for the car to compute. In short, they’re not there yet but are getting close.
5:02am 18 March 2015
We’re now touching on government responsibilities and safety regulations for self-driving cars. Also, the terrifying potential for autonomous cars to get hacked — especially if there’s no steering wheel or brake pedals.
“It’s easy to hack something cosmetic but much harder to hack something that’s dangerous — there will be multiple levels of security,” Musk said.
It’s all conjecture at the moment, but still interesting stuff.
5:20am 18 March 2015
And that’s it! In summary, we got Titan X pricing, some new Pascal GPU details, a Digits Devbox announcement and a general overview of Nvidia’s achievements in deeper learning and self-driving cars. What did you think of the product announcements? Anything tickle your fancy? Have at it in the comments section below!
Gizmodo travelled to GTC 2015 in San Jose, California as a guest of Nvidia.