The NVIDIA GTC 2016 Keynote Live Blog

02:14PM EDT - Off to see more of the show

02:14PM EDT - And that's a wrap

02:13PM EDT - Recapping: SDKs, IRAY VR, Tesla P100, DGX-1, and autonomous cars

02:12PM EDT - Part of the 2016/2017 Formula E season

02:12PM EDT - Wil; be participating in the Roborace. All cars are PX2-powered

02:12PM EDT - Autonomous race car. 2200lbs

02:10PM EDT - Demonstrating DAVENET AI driivng software in action

02:10PM EDT - Cue "Knight Rider" theme

02:09PM EDT - It took BB8 some time to get halfway-decent at driving

02:08PM EDT - Roll video

02:08PM EDT - So we're going to see BB-8 learn to drive

02:08PM EDT - "Name of the car" even

02:08PM EDT - "We've been working on a project that is really fun. The name of the card is BB-8"

02:05PM EDT - PX2 in the car, DGX-1 in the cloud

02:05PM EDT - (Jen-Hsun is prepared for zoom photos this time)

02:04PM EDT - Drive PX2 uses two unannounced Pascal GPUs

02:04PM EDT - Drive PX2 in Jen-Hsun's hands

02:04PM EDT - Baidu, even

02:03PM EDT - Baisu is working on an NVIDIA-powered self-driving car computer as well

02:02PM EDT - Demoing DriveNet running at 180fps running on the smallest Drive PX

01:58PM EDT - Recapping Drive PX

01:57PM EDT - Up now: cars

01:57PM EDT - Deep learning everywhere

01:56PM EDT - Recap: Tesla M40 for hyperscale, K80 for multi-app HPC, P100 for scales very high, and DGX-1 for the early adopters

01:54PM EDT - First DGX-1s will be going to research universities

01:54PM EDT - NVIDIA is taking orders starting today

01:53PM EDT - DGX-1: $129,000

01:50PM EDT - NVIDIA has adapted TensorFlow for VGX-1

01:47PM EDT - Now on stage: Raja Monga of Google's TensorFlow team

01:47PM EDT - More AI/neural network examples coming up

01:41PM EDT - Baidu is using recurrent neural networks rather than convolutional

01:40PM EDT - Now on stage: Brian of Catanzaro of Baidu

01:38PM EDT - 1.33B images per day

01:38PM EDT - "We achieved a 12x speed-up year-over-year" in deep learning

01:37PM EDT - "Datacenter in a box"

01:37PM EDT - Discussing the challenges in scaling out the number of nodes in many algorithms

01:35PM EDT - Two Xeons, and 7TB in SSD capacity

01:35PM EDT - Quad Infiniband, Dual 10GBe

01:34PM EDT - 170TF FP16 in a box. 8 P100s in a hybrid cube mesh

01:34PM EDT - A full deep learning rackmount server

01:34PM EDT - NVIDIA DGX-1

01:33PM EDT - But if it's 600mm2 for just the die, that's a huge jump in the size of dies being produced on 16nm/14nm TSMC/Samsung FinFET

01:33PM EDT - Need to get confirmation on whether 600mm2 is just the GPU die, or if they're counting other parts as well

01:32PM EDT - "We'll ship it... soon"

01:32PM EDT - P100 servers coming in Q'17

01:31PM EDT - P100 in volume production today

01:30PM EDT - NV wanted new algorithms to take advantage of the hardware

01:29PM EDT - Recapping NVLink. 5x the aggregate speed of PCIe 3.0

01:29PM EDT - "TSMC CoWoS® (Chip-On-Wafer-On-Substrate) services use Through Silicon Via (TSV) technology to integrate multiple chips into a single device. This architecture provides higher density interconnects, decreases global interconnect length, and lightens associated RC loading resulting in enhanced performance and reduced power consumption on a smaller form factor."

01:28PM EDT - Chip on Wafer on Substrate, the largest such chip ever made

01:28PM EDT - Er, 600mm^2 !!!!

01:27PM EDT - 500mm^2 !!!

01:27PM EDT - Jen-Hsun is "very frickin excited" about it

01:27PM EDT - Pre-emption

01:27PM EDT - Pascal, 16nm FinFET, Chip-On-Wafer-On-Substrate, NVLink, and New AI Algorithms

01:26PM EDT - The Tesla P100 is "5 miracles"

01:26PM EDT - (150B Transistors is undoubtedly counting the RAM, BTW)

01:26PM EDT - This is using the previously announced mezzanine connector with on-package memory

01:25PM EDT - (14MB is huge for a register file, BTW. That's a lot of very fast memory)

01:24PM EDT - 5.3TF FP64, 10.6TF FP32, 21.2TF FP16, 14MB SM Register File, 4MB L2 Cache

01:24PM EDT - 150B Transistors

01:23PM EDT - "The most ambitious project we have ever undertaken"

01:23PM EDT - Tesla P100

01:23PM EDT - writeln ('Hello, world.')

01:23PM EDT - Pascal time!

01:22PM EDT - AI needs more computing power than what is currently available

01:22PM EDT - "We simply don't have enough computing horsepower"

01:18PM EDT - Teaching AI to draw landscapes inspired by those images

01:18PM EDT - Training it with Romantic-era images

01:17PM EDT - Teaching a neural network to paint

01:17PM EDT - Demo time: Facebook AI Research

01:16PM EDT - Jen-Hsun wants to move from supervised, labor-intensive learning to unsupervised learning

01:14PM EDT - GIE: 20 images/s/W on the Tesla M4

01:14PM EDT - "There's no reason to use FPGAs. There's no reason to design dedicated chips"

01:13PM EDT - (Maxwell powered Tesla cards: http://www.anandtech.com/show/9776/nvidia-announces-tesla-m40-m4-server-cards-data-center-machine-learning )

01:13PM EDT - Recapping the Tesla M40 and M4

01:13PM EDT - Hyperscale is NVIDIA's fastest growing market

01:11PM EDT - Achieving super-human results without super-humans to program them

01:10PM EDT - Side note: bits of this remind me of the hard AI era of the 80s, when at one point true AI was thought to be right around the corner

01:10PM EDT - Talking about how increasingly broad companies are dipping their toes in AI and deep learning

01:08PM EDT - "Cloud platforms of the future are going to be powered by AI"

01:06PM EDT - Jen-Hsun is recapping areas where deep learning has ultimately come up with better algorithms than human-created programs

01:03PM EDT - "Deep learning is a big deal"

01:03PM EDT - "You've heard me talk about deep learning for for the last five years"

12:59PM EDT - Deep learning is key

12:57PM EDT - Microsoft ImageNet has been able to beat a human at image recognition

12:57PM EDT - This year will mark a major year in AI

12:55PM EDT - Up next: AI

12:54PM EDT - Not as capable as Iray VR, but it can handle a single photosphere

12:54PM EDT - Also announcing Iray VR Lite

12:53PM EDT - Jen-Hsun wants to get to the point where he can get out of his car at the office and it'll go park itself underground

12:51PM EDT - (This is on an HTC Vive, for anyone keeping track)

12:51PM EDT - Iray VR rendering of the inside of NVIDIA's new, under-construction headquarters

12:49PM EDT - Using probes to mark out fixed locations. Each probe takes 1 hour on an 8 GPU Quadro setup

12:48PM EDT - Raytracing render for VR applications

12:48PM EDT - Now announcing Iray VR

12:48PM EDT - The Mars demo was running on a Titan, but Jen-Hsun believes that's not enough. Need more performance to better physically simu;ate light

12:46PM EDT - Vendors have been pushing VR as an experience, and this is one such idea

12:45PM EDT - Geeks on Mars

12:44PM EDT - Woz is the first person to try Mars 2030 in VR

12:43PM EDT - Woz would seriously go if it were possible

12:42PM EDT - Jen-Hsun wants to make the Woz the first person on Mars

12:41PM EDT - Steve Wozniak has called in

12:40PM EDT - Demoing Mars 2030 live (though not in VR)

12:37PM EDT - Also in the VR tour, Mars 2030, an 8km^2 reconstruction of the surface of Mars

12:35PM EDT - Roll video

12:34PM EDT - Everest VR will be demoed at the show's VR area

12:33PM EDT - How will VR transform communications, design, and more?

12:32PM EDT - Video games are a very clear use of VR. But what about other fields?

12:31PM EDT - "A brand new computing platform"

12:31PM EDT - Up next: VR

12:30PM EDT - TX1 can process 24 images per second per Watt

12:29PM EDT - GIE is specifically for inference, as opposed to training on cuDNN

12:29PM EDT - NVIDIA wants to boost overall image processing throughput and energy efficiency

12:28PM EDT - New platform: GIE, the GPU Inference Engine

12:27PM EDT - Finally, Jetpack: The Jetson/Tegra X1 software ecosystem

12:27PM EDT - Close partners (JPL partners) have early access to the current testing builds

12:26PM EDT - General release in Q1'17

12:26PM EDT - DriveWorks: SensorFusion, computer vision/detection

12:25PM EDT - VRWorks: VR SLI, Context Priority, Multi-Res Shading

12:24PM EDT - CUDA 8 confirmed to support Pascal

12:23PM EDT - If CUDA 8 is due in June, it stands to reason we may be seeing Pascal around that time in some form

12:22PM EDT - CUDA 8 available in June

12:22PM EDT - ComputeWorks: cuDNN, nvGRAPH, IndeX

12:20PM EDT - DesignWorks: MDL libraries, Optix, etc

12:19PM EDT - Recapping the latest GameWorks features such as voxel accelerated ambient occlusion

12:18PM EDT - This seems to be a bundling of the various NVIDIA SDKs, including GameWorks, DriveWorks, and VRWorks

12:17PM EDT - First up: announcing the NVIDIA SDK

12:17PM EDT - Jen-Hsun will be talking about 5 things: toolbox, deep learning chip, deep learning software, VR, & more

12:15PM EDT - Over 300,000 registered CUDA developers

12:15PM EDT - This despite the fact that they now hold multiple GTCs over the globe

12:15PM EDT - GTC is getting bigger than ever. Twice as large as GTC 2012

12:13PM EDT - Discussing that they do all of this for "you", the audience, and its vast computing needs

12:12PM EDT - Jen-Hsun is now on stage

12:11PM EDT - Self-driving cars, Go, and more

12:11PM EDT - The video theme: AI

12:09PM EDT - Roll video

12:09PM EDT - Lights are dimming. It's showtime

12:07PM EDT - The keynote hall is at capacity, so it's a full house this morning

12:07PM EDT - Everyone is being asked to take their seat

12:05PM EDT - Meanwhile I haven't seen any sign of a car yet, but it would be typical for Jen-Hsun to work a car into his presentation somehow

12:04PM EDT - NVIDIA is indeed running a few minutes late; it sounds like we may start at 9:10 or so

12:04PM EDT - WiFi is being a bit finnicky, but hopefully we'll be okay

12:00PM EDT - Meanwhile someone behind me is discussing the weather. Yesterday it was in the 70s; by tomorrow it's in the 90s. The response from the person next to them: boy, those GPUs sure are hot!

11:59AM EDT - OpenPOWER partners will be holding a keynote tomorrow to discucss the latest advancements in that platform. For NVIDIA it's a big deal as OpenPOWER was previously setup to support NVLink between the CPU and GPUs

11:58AM EDT - On a side note: also taking place in San Jose at the convention center is the show-within-a-show OpenPOWER conference

11:57AM EDT - We may see some consumer news as well, but that's a big if at this point

11:57AM EDT - Some pro visualization (Quadro) news is also likely, given the heavy focus on VR here by the exhibitors

11:57AM EDT - At a minimum I expect Tesla/HPC news on this front, as this is what NVIDIA has focused on in the past, and the HPC market isn't as timing-sensitive

11:56AM EDT - The big news this year will of course be NVIDIA's Pascal architecture, which is scheduled to launch this year

11:56AM EDT - The keynote is scheduled to start at 9am PT, however I suspect we're going to be a few minutes late

11:53AM EDT - Kicking things off as always is the NVIDIA keynote, presented by CEO Jen-Hsun Huang

11:52AM EDT - We're here in sunny San Jose for the 2016 edition of NVIDIA's annual GPU Technology Conference (GTC)