
Original Link: https://www.anandtech.com/show/12861/computex-2018-nvidia-live-blog-1pm-tw-5am-utc
Computex 2018 NVIDIA Live Blog (1pm TW, 5am UTC)
by Ian Cutress on June 4, 2018 12:47 AM EST- Posted in
- GPUs
- Trade Shows
- NVIDIA
- Live Blog
- Computex 2018

12:51AM EDT - We're here at NVIDIA for our first Live Blog of Computex. Event starts in a few minutes.
12:52AM EDT - Surprisingly everything on stage is not under wraps.
12:52AM EDT - One wonders what NVIDIA will be announcing
12:52AM EDT - Jensen approached the press crowd outside the event with sandwiches and cookies for lunch. Easy PR move
12:54AM EDT - Disclosure: I haven't had lunch. The sandwich was green. I haven't had time to eat it yet.
12:55AM EDT - Up on the stage is an HGX-2 (OEM version of DGX-2), a V100, Xavier
12:56AM EDT - Isn't Xavier meant to have two next-gen GPUs? It looks like it has the older GPUs on for now
12:56AM EDT - Sorry, Pegasus
12:58AM EDT - Looks like the V100 being used
12:58AM EDT - OK here we start
12:59AM EDT - Jensen is handing out more cookies
12:59AM EDT - Last week we had GTC Taiwan
12:59AM EDT - GTC became world's largest developer conference
01:00AM EDT - More GTC spin-out events
01:00AM EDT - Japan, China, Europe, Israel, Washington DC etc
01:00AM EDT - We know the future of computing is going to look different
01:00AM EDT - First update is on GeForce
01:00AM EDT - Technology demand for the processors is growing exponentially
01:01AM EDT - Second topic is AI, Third is AI implications, Fourth is a new market NVIDIA is pursing
01:01AM EDT - Last year went for thin form factor GPUs with MaxQ
01:01AM EDT - Gaming notebooks at 18 mm, 5 lbs, GTX 1080
01:04AM EDT - The market wants thin and light notebooks
01:04AM EDT - Last year we had 8 MaxQ. Now we have 26
01:05AM EDT - PC Gaming is doing great and is vibrant
01:05AM EDT - Always new things you can do with it
01:05AM EDT - New content continues to get richer
01:05AM EDT - always reinventing itself
01:06AM EDT - In 8 months, PC gaming market increased due to Battle Royale gaming type
01:06AM EDT - +10M weekly gamers in 8 months
01:06AM EDT - Production value of games also increasing
01:06AM EDT - >Gaming is bigger than film
01:07AM EDT - Game consoles create a positive feedback system
01:07AM EDT - Developers have a larger install base for GPUs and gaming
01:07AM EDT - GTX class GPUs went from 17% of NV to 50% in 5 years
01:08AM EDT - This fall will see three games: COD Black Ops 4 (with Battle Royale mode), Battlefield 5, Shadow of the Tomb Raider
01:08AM EDT - 'This year will be pretty amazing for GeForce'
01:09AM EDT - New era of computing coming
01:09AM EDT - Two powerful forces coming to change computing architecture
01:10AM EDT - How computers are made has to change - how software will be used has to be incredible
01:10AM EDT - >Bifurcation of single thread vs multi
01:10AM EDT - More demand for computing horsepower than before
01:10AM EDT - DL and AI driving more compute than ever
01:11AM EDT - In time, software will write software
01:11AM EDT - "What kind of computer do you have to build to enable software to write software?"
01:12AM EDT - We are expecting application demand to increase over 1000 in 10 years
01:12AM EDT - We have to advance computing performance by at least the same scale - we do it with GPU computing
01:12AM EDT - Today, talking about 3 phases talking about this new computing model
01:12AM EDT - Phase 1 is for the industry to create the fundamental computing platform, hardware and software, and the tools
01:13AM EDT - Create the ecosystem to take systems over the workd
01:13AM EDT - Now showing DGX-2
01:13AM EDT - Announced before. 16x V100 GPUs
01:13AM EDT - V100 fuses computational with high precision numerics and deep learning
01:14AM EDT - Need to compute at a large dynamic range
01:14AM EDT - Volta Tensor Cores
01:14AM EDT - Volta is 21 B transistors at 12FF
01:14AM EDT - 32GB of HBM
01:14AM EDT - 125 TFLOPS of Tensor Core Perf in one chip
01:14AM EDT - 2 PFLOPS in one box
01:15AM EDT - Parallel computing in two ways on DGX-2
01:15AM EDT - Single model per GPU, or One model across system
01:15AM EDT - Scale out vs Scale up
01:15AM EDT - Scale up is more difficult
01:15AM EDT - Teaming 16 GPUs to do a single model
01:15AM EDT - Requires high levels of communication
01:16AM EDT - Requires software that can do this
01:16AM EDT - All 16 GPUs can communicate at the same time
01:16AM EDT - Same model to compute one vs 16 with DGX-2
01:17AM EDT - 300 servers replaced with one node
01:17AM EDT - Lots of speed records
01:17AM EDT - Fastest single chip
01:17AM EDT - Fastest single chip
01:18AM EDT - Fastest single node
01:18AM EDT - Fastest at scale
01:18AM EDT - Fastest Inference
01:18AM EDT - Highest Inference Throughput
01:18AM EDT - Also DGX-2 becomes HGX-2
01:18AM EDT - Made available for hyperscale datacenters
01:19AM EDT - Six switches at the bottom for the 8 GPUs
01:19AM EDT - at the bottom
01:20AM EDT - >I can't see USB ports
01:20AM EDT - Form factor is as important as the ATX form factor was for desktop computing
01:20AM EDT - "Hyperscale computing element"
01:22AM EDT - Different architectures for different computing models
01:22AM EDT - Architectures of all sizes and shapes
01:23AM EDT - HGX-T1, HGX-T2, HGX-I1, HGX-I2, SCX-E1, SCX-E2, SCX-E3, SCX-E4
01:23AM EDT - All worlds leading server makers support NVIDIA
01:23AM EDT - AI revolutionizing large markets
01:23AM EDT - Hardware, software, deployment, ecosystem
01:24AM EDT - Focused on six large industries for large GPU servers
01:24AM EDT - Supercomputing (CUDA), AI, Drive, NVIDIA Clara, NVIDIA Metropolis, NVIDIA RTX
01:25AM EDT - Clara is a computing system with a software stack for medical imaging, construction, and detection
01:25AM EDT - Taking RAW data for AI implementation in medical
01:26AM EDT - Metropolis is a software plaform that allows billions of images from live video to detect areas of concern
01:26AM EDT - >I assume that's the same as tracking people/vehicles across a city. How the city uses that is up to them
01:27AM EDT - RTX - rasterization with real-time ray tracing with AI prediction
01:27AM EDT - RTX is basically the future of gaming as well
01:27AM EDT - bringing in ray tracing 5-10 years earlier than expected
01:30AM EDT - Now doing a real-time demo
01:31AM EDT - >Everyone's phone is going off because Taiwan is doing an air raid drill
01:32AM EDT - Deep learning with ray tracing is the way forward
01:32AM EDT - This demo requires four volta GPUs
01:34AM EDT - Third phase is about AI to be able to come out of the computer into the world
01:34AM EDT - In order to do this, need three things: rich sensors, high perf/low power AI brain, smart actuators
01:35AM EDT - NVIDIA focuses on the brain
01:35AM EDT - enabling intelligent machiens
01:35AM EDT - We used to had to go out to get food. Now food comes to us
01:36AM EDT - Delivery, Industrial Drones, Agriculture, Pick-and-place, Logistics, Manufacturing
01:36AM EDT - Xavier, 9B transistors
01:37AM EDT - 8-core ARM CPU, 512 CUDA Cores
01:37AM EDT - Designed for ASIL-D
01:38AM EDT - New deep learning accelerator: DLA
01:38AM EDT - DLA is NVIDIA's version of the TPU
01:39AM EDT - Working on self-driving car
01:40AM EDT - Different versions of hardware for different levels of compute
01:40AM EDT - Drive Pegasus with dual Xavier and dual Volta
01:40AM EDT - (so not post-Volta?)
01:40AM EDT - Pegasus for driverless taxis
01:42AM EDT - Standard self-driving video time
01:44AM EDT - Announcing the Jetson Xavier Devkit $1299
01:45AM EDT - Computing platform
01:45AM EDT - 12nm FF
01:46AM EDT - Single largest processor project in NV history
01:46AM EDT - Production silicon is here
01:46AM EDT - Comes in small module
01:47AM EDT - Small dev kit
01:47AM EDT - 8 high-end ARM cores, stereo vision, 30W
01:48AM EDT - Jetson Xavier, early access in August
01:49AM EDT - On top of Jetson Xavier, have DL applications
01:50AM EDT - ISAAC IMX (Intelligent Machine Acceleration Applications)
01:50AM EDT - POSE estimation - infer the pose
01:50AM EDT - Gesture recognition
01:50AM EDT - Facial recognition and tracking
01:50AM EDT - Speech
01:51AM EDT - Speech recognition on device. Latency is key
01:51AM EDT - Eye tracking
01:51AM EDT - Hand Pose and finger skeletal tracking
01:52AM EDT - Environment depth
01:52AM EDT - DL to learn depth
01:52AM EDT - Visual Odometry
01:53AM EDT - Teaching robots in virtual reality in the server - test the hardware in the simulation
01:54AM EDT - Put the hardware and software in to the robot and it should just work
01:55AM EDT - ISAAC is the third phase of the AI revolution
01:56AM EDT - Apply intelligent machines to drive every industry
01:56AM EDT - Showing contruction and agriculture uses on video
01:56AM EDT - The two typical go-to demos
01:58AM EDT - Flying cameras/drones with AI
01:58AM EDT - NVIDIA's ' i am ai ' campaign
01:59AM EDT - Taiwan research into AI collaboration
01:59AM EDT - Focusing on manufacturing, healthcare, smart city and transportation
01:59AM EDT - 10-year collab
01:59AM EDT - work with the government and research institutions
01:59AM EDT - Building a supercomputing infrastructure
02:00AM EDT - Build a v.fast supercomputer
02:00AM EDT - self-driving car platforms
02:04AM EDT - Software written on HGX-2 can scale down to the Jetson Xavier
02:05AM EDT - First Gen Pegasus is Volta, Second Gen with Post-Volta coming
02:07AM EDT - Q: When is Volta coming to gaming? A: So many gamers were deprived of GeForces due to mining, but now pricing is coming down. 1080/Ti are the best cards right now, we suggest they buy GeForce ASAP
02:13AM EDT - Q: How long to AI accelerating ray tracing for games? Next gen or beyond? A: It's going to take time. Tackling film and enterprise/commercial first, then some day to mass market.
02:15AM EDT - Q: When can we run 4K 120 Hz on a single card. A: I don't know ... (!) I could predict, but I'm not going to
02:15AM EDT - Q: Mining only cards? A: We do these already to diffuse the demand for GeForce, Mining is a bonus to our business, but it is not our business
02:17AM EDT - Q: When is the next GeForce? A: I'll invite you. There will be lunch. But it's a long time.
02:20AM EDT - Q: Anything to say about NVIDIA's talk at Hot Chips about Next Gen changing to TBD? A: Live in the present! We will invite you to our launch events
02:22AM EDT - Q: Competitors combine CPU + GPU. What can NV do? A: Xavier is already CPU + GPU + other DSPs. Integration works great in certain markets. We also have DGX-2 when focus on GPUs, we have Xavier for unified platforms
02:29AM EDT - I think that's a wrap.