GTC Europe 2016: NVIDIA Keynote Live Blog with CEO Jen-Hsun Huangby Ian Cutress on September 28, 2016 4:17 AM EST
04:24AM EDT - I'm here at the first GTC Europe event, ready to go for the Keynote talk hosted by CEO Jen-Hsun Huang.
04:25AM EDT - This is a satellite event to the main GTC in San Francisco. By comparison the main GTC has 5000 attendees, this one has 1600-ish
04:25AM EDT - This is essentially GTC on the road - they're doing 5 or 6 of these satellite events around the world after the main GTC
04:27AM EDT - We're about to start
04:29AM EDT - Opening video
04:30AM EDT - 'Deep Learning is helping farmers analyze crop data in days what used to take years'
04:30AM EDT - 'Using AI to deliver relief in harsh conditions' (drones)
04:30AM EDT - 'Using AI to sort trash'
04:30AM EDT - Mentioning AlphaGO
04:31AM EDT - JSH to the stage
04:32AM EDT - 'GPUs can do what normal computing cannot'
04:32AM EDT - 'We're at the beginning of something important, the 4th industrial revolution'
04:33AM EDT - 'Several things at once came together to make the PC era something special'
04:33AM EDT - 'In 2006, the mobile revolution and Amazon AWS happened'
04:33AM EDT - 'We could put high performance compute technology in the hands of 3 billion people'
04:34AM EDT - '10 years later, we have the AI revolution'
04:34AM EDT - 'Now, we have software that writes software. Machines learn. And soon, machines will build machines.'
04:35AM EDT - 'In each era of computing, a new computing platform emerged'
04:35AM EDT - 'Windows, ARM, Android'
04:35AM EDT - 'A brand new type of processor is needed for this revolution - it happened in 2012 with the Titan X'
04:37AM EDT - 'Deep Learning was in the process, and the ability to generalize learning was a great thing, but it had a handicap'
04:37AM EDT - 'It required a large amount of data to write its own software, which is computationally exhausting'
04:38AM EDT - 'The handicap lasted two decades'
04:38AM EDT - 'Deep Neural Nets were then developed on GPUs to solve this'
04:39AM EDT - 'ImageNet Classification with Deep Convolutional Neural Networks' by Alex Krizhevsky at the University of Toronto
04:40AM EDT - 'The neural network out of that paper, 'AlexNet' beat seasoned Computer Vision veterans with hand tuned algorithms at ImageNet'
04:40AM EDT - 'One of the most exciting events in computing for the last 25 years'
04:41AM EDT - 'Now, not a week goes by when there's a new breakthrough or milestone reached'
04:42AM EDT - 'e.g., 2015 where Deep Learning beat humans at ImageNet, 2016 where speech recognition reaches sub-3% in conversational speech'
04:43AM EDT - 'As we grow, the computational complexity of these networks becomes even greater'
04:44AM EDT - 'Now, Deep Learning can beat humans at image recognition - it has achieved 'Super Human' levels'
04:44AM EDT - 'One of the big challenges is autonomous vehicles'
04:44AM EDT - 'Traditional CV approaches wouldn't ever work for auto'
04:45AM EDT - 'Speech recognition is one of the most researched areas in AI'
04:45AM EDT - 'Speech will not only change how we interact with computers, but what computers can do'
04:47AM EDT - Correction, Microsoft hit 6.3% error rate in speech recognition
04:47AM EDT - 'The English language is fairly difficult for computers to understand, especially in a noisy environment'
04:48AM EDT - 'Reminder, humans don't achieve 0% error rate'
04:48AM EDT - 'These three achievements are great: we now have the ability to simulate human brains: learning, sight and sound'
04:49AM EDT - 'The ability to perceive and the ability to learn are fundamentals of AI - we now have the three pillars to solve large-scale AI problems'
04:49AM EDT - 'NVIDIA invented the GPU, and 10 years ago we invented GPU computing'
04:50AM EDT - 'Almost all new high performance computers are accelerated, and NVIDIA is in 70% of them'
04:50AM EDT - 'Virtual Reality is essentially computing human imagination'
04:51AM EDT - 'Some people have called NVIDIA the AI Computing Company'
04:51AM EDT - JSH: 'We're still the fun computing company, solving problems, and most of the work we do is exciting for the future'
04:52AM EDT - 'Merging simulation, VR, AR, and powered by AI, and scenes like Tony Stark in Iron Man captures what NVIDIA is going after'
04:53AM EDT - GTC 2014 to 2016: 4x attendees, 3x GPU developers, 25x deep learning devs, moving from 2 events/year to 7 events/year worldwide
04:54AM EDT - The GPU devs and Deep Learning devs numbers are 'industry metrics', not GTC attendees
04:55AM EDT - 'I want the developers to think about the Eiffel Tower - an iconic image in Europe'
04:56AM EDT - 'The brain typically imagines an image of the tower - your brain did the graphics'
04:56AM EDT - 'The brain thinks like a GPU, and the GPU is like a brain'
04:57AM EDT - 'The computer industry has invested trillions of dollars into this'
04:57AM EDT - 'The largest supercomputer has 16-18000 NVIDIA Tesla GPUs, over 25m CUDA cores'
04:58AM EDT - 'GPUs are at the forefront of this'
04:58AM EDT - 'GPU Deep Learning is actually a new model'
04:59AM EDT - 'Previously, software engineers write the software, QA engineers test it, and in production the software does what we expect'
04:59AM EDT - 'GPU Deep Learning is a bit different'
04:59AM EDT - 'Learning is important - a deep neural net has to gather data and learn from it'
05:00AM EDT - 'This is the computationally intensive part of Deep Learning'
05:00AM EDT - 'Then the devices infer, using the generated neural net'
05:00AM EDT - 'GPUs have enabled larger and deeper neural networks that are better trained in shorter times'
05:01AM EDT - 'A modern neural network has hundreds of hidden layers and learns a hierarchy of features'
05:01AM EDT - 'Our brain has the ability to do that. Now we can do that in a computer context'
05:02AM EDT - 'The trained network is placed into data centers for cloud inferencing with large libraries to answer questions from its database'
05:02AM EDT - 'This is going to be big in the future. Every question in the future will be routed through an AI network'
05:03AM EDT - 'GPU inferencing makes response times a lot faster'
05:03AM EDT - 'This is the area of the intelligent device'
05:04AM EDT - 'This is the technology for IoT'
05:04AM EDT - 'AI is comparatively small coding with lots and lots of computation'
05:05AM EDT - 'The important factor of neural nets is that they work better with larger datasets and more computation'
05:05AM EDT - 'It's about the higher quality network'
05:05AM EDT - 2012 AlexNet was 8-layers, 1.4 GFlops, 16% error
05:06AM EDT - 2015 ResNet is 152 Layers, 22.6 GFlops for 3.5% error
05:06AM EDT - The 2016 winner has improved this with a network 4x deeper
05:06AM EDT - Baidu in 2015, using 12k hours of data and 465 GFLOPs can do 5% speech recognition error
05:07AM EDT - 'This requires a company to push hardware development at a faster pave than Moore's Law'
05:07AM EDT - 'So NVIDIA thought, why not us'
05:08AM EDT - 'I want to dedicate my next 40 years to this endeavour'
05:08AM EDT - 'The rate of change for deep learning has to grow, not diminish'
05:09AM EDT - 'The first customer of Pascal is an open lab called OpenAI, and their mission is to democratize the AI field'
05:10AM EDT - 'Our platform is so accessible - from a gaming PC to the cloud, supercomputer, or DGX-1'
05:11AM EDT - 'You can buy it, rent it, anywhere in the world - if you are an AI researcher, this is your platform'
05:11AM EDT - 'We can't slow down against Moore's Law, we have to hypercharge it'
05:12AM EDT - 'We also have partners, such as IBM with cognitive computing services'
05:13AM EDT - 'We worked with IBM to develop NVLink to pair POWER8 with the NVIDIA P100 GPU to get a network of fast processors and fast GPUs, dedicated to solve AI problems'
05:13AM EDT - 'Today we are announcing a new partner'
05:13AM EDT - 'To apply AI for other companies worldwide'
05:14AM EDT - New partner is SAP
05:14AM EDT - 'We want to bring AI to companies around the world as one of the biggest AI hardware/software collaborations'
05:14AM EDT - 'We want SAP to be able to turbocharge their customers with Deep Learning'
05:15AM EDT - 'The research budget of DGX-1 was $2b, with 10k engineer years of work'
05:15AM EDT - 'DGX-1 is now up for sale. It's in the hands of lots of high impact research teams'
05:17AM EDT - 'We're announcing two designated research centers for AI research in Europe - one in Germany, one in Switzerland, with access to DGX-1 hardware'
05:17AM EDT - Now Datacenter Inferencing
05:18AM EDT - 'After the months and months of training of the largest networks, when the network is complete, it requires a hyperscale datacenter'
05:18AM EDT - 'The market goes into 10s of millions of hyperscale servers'
05:18AM EDT - 'With the right inference GPU, we can provide instantaneous solutions when billions of queries are applied at once'
05:19AM EDT - 'NVIDIA can allow datacenters to support a factor million more load without ballooning costs or power a million times'
05:20AM EDT - 'NVIDIA recently launched the Tesla P4 and P40 inferencing accelerators'
05:20AM EDT - 'You can replace 3-4 racks with one GPU'
05:21AM EDT - Here's the link to our write-up of the P4/P40 news: http://www.anandtech.com/show/10675/nvidia-announces-tesla-p40-tesla-p4
05:21AM EDT - P4 at 50W, P40 at 250W, using Pascal and using INT8
05:22AM EDT - Announcing TensorRT
05:22AM EDT - 'Software package that compiles, fuses operations and autotuning, optimizing the code to the GPU for efficiency'
05:22AM EDT - 'Support a number of networks today, plan to support all major networks in time'
05:23AM EDT - 'Live video is the increasingly shared content of importance'
05:24AM EDT - 'It would be nice if, as you are uploading the live video, it would identify which of your viewers/family would be interested'
05:24AM EDT - 'AI should be able to do this'
05:24AM EDT - Example on stage of 90 streams at 720p30 running at once
05:25AM EDT - A human can do this relatively easily, but a bit slowly
05:25AM EDT - 'A computer can learn from the videos what is happening and assign relevant tags'
05:26AM EDT - 'The future needs the ability to filter based on tags generated from AI live video as it is being generated'
05:28AM EDT - 'We trained a network to determine the style of various paintings, and repaint it in a different style'
05:29AM EDT - 'We want to be able to do it live'
05:29AM EDT - redrawing every single frame
05:30AM EDT - 'The neural network generates new art'
05:30AM EDT - 'frame by frame'
05:31AM EDT - 'Applications for deep learning are clearly very broad'
05:32AM EDT - 'Today, 2000 active deep learning implementations are in use for customers'
05:33AM EDT - 'Partners are ready to configure servers with Deep Learning embedded'
05:35AM EDT - 'ODMs and Server Builders can configure a Deep Learning system for any customer'
05:36AM EDT - Deep Learning for advertising is interesting. 'Adding ads to live video based on viewer preferences'
05:37AM EDT - 'AI Startups or startups using deep learning are everywhere, in lots of diverse fields'
05:38AM EDT - 'Benevolent.ai is the first Europe customer of DGX-1'
05:38AM EDT - 'They can take what used to take a year, and now complete it in a week or two'
05:40AM EDT - 'not all faces are straight on to the camera, so AI can tell if things have changed (hair, age, emotion)'
05:40AM EDT - 'Deep learning can be applied for intelligent voice, to detect if someone is lying on an insurance claim or similar'
05:40AM EDT - 'Deep Learning can detect plastic in waste that is suitable for recycling'
05:41AM EDT - Now inferencing for devices
05:41AM EDT - 'Intelligent Devices, or what people call IoT'
05:42AM EDT - 'Only our imagination limits what the devices are'
05:42AM EDT - >This is the usual thing for companies focusing on AI hardware - leave the 'killer device' problem to others and just provide the hardware during development.
05:43AM EDT - 'Internet connected, intelligent machines, that have the capability of AI, are the future'
05:44AM EDT - Mentioning Jetson TX1, >20 images/sec/W in a 10W platform (adjust as required)
05:45AM EDT - 'We're starting an institute for applying deep learning - the NVIDIA Deep Learning Institute'
05:45AM EDT - 'NVIDIA DLI places always sell out, so we're expanding the reach'
05:46AM EDT - 'AI Transportation is a $10T industry: safety, accessibility and ease of use'
05:46AM EDT - 'Autonomous vehicles is not just about smart sensors - it's a compute problem'
05:48AM EDT - 'AI Auto has to perceive, reason, understand, plan, and apply'
05:49AM EDT - 'NVIDIA has jumped in with both feet to create a scalable platform for the Auto industry'
05:49AM EDT - 'Scalability means different segments of the industry have different visions for autonomous vehicles'
05:50AM EDT - Cruise hardware vs. auto chauffeur vs full autonomy
05:51AM EDT - >I just had a thought, it'd be interesting as to what Taxi drivers think about this. If people can just say 'take me home'...
05:51AM EDT - 'Drive PX 2 is perception, reasoning, driving, AI processing, algorithms and software with scalability'
05:52AM EDT - 'When we first showed off PX2, it was the large full-fat version as that's what the big customers wanted'
05:53AM EDT - 'Drive PX 2 Auto Cruise connects to two cameras and identifies a HD map, at 10W'
05:54AM EDT - 'It's a HD map all the way to the cloud and supercomputer'
05:54AM EDT - 'Baidu selected Drive PX 2 for their self-driving cars - DriveWorks is connected to the cloud'
05:55AM EDT - 'NVIDIA is announcing a partnership with TomTom as a mapping partner'
05:56AM EDT - 'Tesla for cloud HD map processing, Drive PX 2 for in-car processing, open cloud-to-car platform using AI'
05:56AM EDT - 'Has to be accurate, coherent, and accurate within a few centimenters'
05:56AM EDT - 'We want to continuously crunch data as it is recorded'
05:57AM EDT - Alain De Taeye from TomTom on stage
05:58AM EDT - 'We've mapped 70% of society, 47 million miles'
05:59AM EDT - 'we have basic information, but we need to make accurate HD maps in an affordable way'
06:00AM EDT - 'TomTom gets 7 billion traces a day which can be processed'
06:00AM EDT - 'We're 120k miles into the 60million mile challenge for HD maps'
06:01AM EDT - 'We need to process video from the cars in super-real time using AI'
06:01AM EDT - 'Using the car to maintain the HD map is the holy grail'
06:01AM EDT - 'Self-driving cars requires an accurate HD map'
06:02AM EDT - Announcing Driveworks Alpha 1, an OS for self-driving cars
06:03AM EDT - 'the ultimate real-time supercomputing problem'
06:03AM EDT - ''I tried' is not an acceptable answer for self-driving cars'
06:04AM EDT - 'The OS uses three neural networks for different things'
06:05AM EDT - DriveNet, OpenRoadNet and PilotNet
06:06AM EDT - >Just had a thought. HD maps means identifying individual trees and triangulating position on a road for self-driving based on an internal schematic, and updating it if there's weather etc. That's important
06:06AM EDT - 'Driving is a behaviour, like playing tennis, it's somewhat automatic and by reflex'
06:07AM EDT - 'PilotNet is a behaviour network'
06:08AM EDT - 'The occupancy grid is tested against what PilotNet wants to do, based on everything it has been taught and also future predictions'
06:08AM EDT - 'We need to continually test what we see with what the car sees and how the AI car reacts'
06:09AM EDT - Videos showing a couple of demos
06:09AM EDT - Detecting objects on a front facing vehicle camera
06:10AM EDT - 'Where other cars are, where the lanes are, where it is safe to drive - all done via neural network inferencing'
06:10AM EDT - 'Where is it ok to drive is more important than where you need to drive - it keeps it safe'
06:12AM EDT - 'The occupancy grid combines the sensor data and HD maps to generate it's own 3D map of where the vehicle is'
06:13AM EDT - 'Caffeine is the fuel of deep learning engineers'
06:15AM EDT - 'When we're driving, we don't do calculus, but the AI has to'
06:16AM EDT - 'NVIDIA BB8 is a car without any prior knowledge, the idea of imitating the driver'
06:17AM EDT - 'We don't have to describe equations, the car will generalise driving behaviour by repeated sampling'
06:18AM EDT - 'Will leave the road to remain safe'
06:19AM EDT - 'BB8 learned how to drive like us'
06:20AM EDT - The sparkles are BB8's neurons firing regarding information that it thinks is important to its behaviour
06:20AM EDT - 'It learned to stay in the middle of the lane, in the middle of the road, a road in the dark etc'
06:20AM EDT - 'It learned that we don't drive over bushes'
06:21AM EDT - 'Driveworks is an open platform for OEMs and car companies to pick and choose the bits they want/need for their solutions'
06:22AM EDT - 'today's announcement is that Alpha 1 is being launched to top tier partners in October with updates every two months'
06:22AM EDT - 'NVIDIA has many AI self-driving cars in development with different partners'
06:24AM EDT - Announcing Xavier
06:25AM EDT - AI Supercomputer SoC
06:25AM EDT - 8-core custom ARM CPU, 512-core Vota, dual 8K video processors, New CV accelerator
06:25AM EDT - 7 billion transistors, 16nm FF+
06:25AM EDT - No perf numbers... ?
06:26AM EDT - Doesn't say Denver cores, so next-gen Denver with Volta?
06:26AM EDT - Sampling Xavier in Q4 2017
06:27AM EDT - Xavier does 20 TOPS DL, 160 SPECINT in 20W
06:28AM EDT - It's so far out, over 12 months mind
06:28AM EDT - So, HotChips 2017 might show off some new uArch details, just like Denver2 this year
06:30AM EDT - Ryan thinks 1 TOPS/W is going to be a tough target
06:30AM EDT - Looks like that's a wrap for the keynote. I think there's some Q&A with JHH now for the press. Time to ask questions