The NVIDIA GTC 2021 Keynote Live Blog (Starts at 8:30am PT/15:30 UTC)by Ryan Smith on April 12, 2021 9:00 AM EST
11:27AM EDT - Thank you for joining us for another year of the NVIDIA GTC keynote live blog
11:28AM EDT - Whether physical or virtual, GTC is inevitab...ly a lot of news in a short period of time
11:29AM EDT - NVIDIA's revenues have doubled over less than half a decade, and with that so has the number of business they're in
11:29AM EDT - Graphics, AI, automotive, HPC, and most recently networking
11:30AM EDT - So it's a lot for CEO Jensen Huang to go over in (ideally) less than 2 hours
11:30AM EDT - This year will be no exception. With a whole year to prepare, NVIDIA is firing on all cylinders ahead of the show
11:30AM EDT - And here we go
11:32AM EDT - With the virtual show, this year's keynote is pre-recorded. So it should keep a tight pace. Still, according to YouTube, we're looking at a 1 hour and 48 minute recording
11:33AM EDT - Rolling the intro video. "I am AI"
11:33AM EDT - And here's Jensen
11:34AM EDT - Starting right off the bat talking about AI
11:34AM EDT - "AI and 5G are the ingrediants to kickstart the 4th industrial revolution"
11:35AM EDT - Jensen's talk will be in 4 stacks: graphics and omniverse, data center AI and server hardware, edge AI and EGX 5G, and automotive/DRIVE
11:37AM EDT - "With just a GeForce, every student can have a supercomputer"
11:38AM EDT - Now recapping some of the things that NVIDIA's clients have been doing with their hardware
11:38AM EDT - By headcount, NVIDIA is primarily a software company (seriously), and there is no shortage of major computer science researchers set to give talks at this year's show
11:39AM EDT - "Let's start where NVIDIA started: computer graphics"
11:40AM EDT - Recapping last year's introduction of second-generation RTX (Ampere) hardware
11:40AM EDT - Now rolling some video of some recently-released games and future games in development
11:42AM EDT - Suffice it to say, game graphical quality has only continued to get better over the years
11:42AM EDT - And NVIDIA wants ray tracing to push that further
11:42AM EDT - But games aren't everything. NVIDIA is also focused on productivity use of graphics
11:42AM EDT - NVIDIA's Omniverse technology
11:43AM EDT - Which was first announced a couple of years back, and went into beta testing late last year
11:43AM EDT - Omniverse is essentially a shared group simulation and graphics software package
11:44AM EDT - Omniverse is server-hosted, and any RTX client can plug in to see it. Or even resort to streaming for those devices that can't render it locally
11:46AM EDT - In other words, shared collaboration and design within a single 3D project. All with an emphasis on high quality physics and rendering
11:47AM EDT - One particular focus of Omniverse is "digital twins"; creating a virtual copy of a real-world project/location
11:48AM EDT - This is one of NVIDIA's big pulls for its traditional professional graphics clients, especially in the movie and TV production industry
11:48AM EDT - But also robotics, R&D, and pretty much any other use case you can think of where a shared, real-time interface to a model might be useful
11:49AM EDT - (Oh good, someone remembered the teapot. It's not graphics without a Utah teapot!)
11:50AM EDT - NVIDIA's Isaac robotics platform can interface with Omniverse as well
11:50AM EDT - Which among other things, can be used to train robots using a digital twin of a factory within Omniverse
11:51AM EDT - NVIDIA has even created a digital twin of a BMW factory
11:52AM EDT - BMW is using this as part of their planning processes
11:53AM EDT - Discussing an example of using the model to optimize an assembly line for productivity and safety by quickly adjusting the line and relocating various tools/stations
11:53AM EDT - BMW is also deploying logistics robots that are using isaac
11:55AM EDT - It all looks impressive. Though I am curious what the required investment is with respect to art. Someone has to create all of these models, items, and their surface textures
11:56AM EDT - Omniverse connector SDKs from major software packages are available now, with more on the way
11:56AM EDT - Omniverse will be available for commercial use this summer under enterprise licensing
11:56AM EDT - Now on to data centers
11:57AM EDT - Currently discussing virtualization, and the impact of doing it on CPUs
11:58AM EDT - GPUs generate a lot of cross-datacenter traffic. Deep learning added even more to that
11:58AM EDT - And thus NVIDIA's networking processors, the Data Processing Unit (DPU)
11:59AM EDT - The Bluefield family of DPUs was inherited from Mellanox, and now a core part of NVIDIA's offerings
12:00PM EDT - Bluefield is designed to offload a major part of network functions, including all the processing that goes with them, such as SSL and security analysis
12:00PM EDT - Today NVIDIA is announcing Bluefield 3
12:00PM EDT - 400Gbps network processor with 22 billion transistors
12:01PM EDT - And NVIDIA is already working on Bluefield 4 for 2024, which will be around 64B transistors, and incorporate NVIDIA's AI acceleration technology
12:01PM EDT - "Software will be written by software running on AI computers"
12:02PM EDT - Now segueing into NVIDIA's DGX server hardware
12:03PM EDT - DGX A100 series ranges from a workstation-like DGX Station box, up through DGX A100 servers and DGX SuperPods comprised of many A100 servers
12:03PM EDT - Announcing the DGX Station 320G
12:04PM EDT - 2.5 PFLOPS, 320GB of VRAM, and all in 1500W
12:04PM EDT - This is essentially the most powerful box NVIDIA can build that can safely be plugged into a standard North American 115V/15A circuit
12:05PM EDT - (PANAMAX for workstations, if you will)
12:05PM EDT - NVIDIA is also updating the DGX SuperPod
12:06PM EDT - The latest generation SuperPod has added Bluefield 2 DPUs
12:06PM EDT - The 80GB A100, first announced last year, is also an option
12:06PM EDT - Pricing starts at 7 million dollars and scales to 60 million depending on the size of the system
12:06PM EDT - Now on to the next subject: transformers
12:07PM EDT - Natural language transformer machine learning models
12:07PM EDT - "We expect to see multi-trillion parameter models by next year"
12:08PM EDT - Transformer models are growing quickly. The bigger the model, generally the better and more nuanced the results
12:08PM EDT - So NVIDIA has developed their own transformer technology: Megatorn
12:09PM EDT - Announcing the Megatron Triton DGX server
12:09PM EDT - Able to repond to up to 16 simultaneous queries in an instant
12:10PM EDT - Now on to NVIDIA's Clara library of machine learning models and technology for medical research
12:10PM EDT - NVIDIA is adding 4 new models to the Clara Discovery library
12:12PM EDT - Among other tasks, one of the new models can be used to recognize DNA sequences
12:12PM EDT - Meanwhile, Jensen is also pitching NVIDIA's hardware and software for drug discovery
12:13PM EDT - And if that's not enough, how about quantum physics simulations running on GPUs? IBM's doing it
12:14PM EDT - Er, excuse me, quantum computing, not quantum physics
12:15PM EDT - NVIDIA is announcing a new software package, cuQuantum, to help research and simulate quantum computers
12:15PM EDT - cuQuantum is optimized to run on NVIDIA's DGX hardware
12:16PM EDT - Jensen wants cuQuantum to do what cuDNN did for deep learning
12:16PM EDT - Now on to data center server architectures
12:17PM EDT - "Processing large amounts of data remains a challenge for computers today"
12:18PM EDT - Discussing the current architecture of GPU server boxes like NVIDIA's DGX: 4 GPUs hooked up to a single CPU via PCI Express
12:18PM EDT - PCI Express is the bottleneck
12:18PM EDT - NVIDIA has NVLink, but no x86 CPU has NVLink
12:18PM EDT - So NVIDIA is making their own data center CPU: Grace
12:18PM EDT - Named after Grace Hopper
12:19PM EDT - Grace is an Arm-based CPU, specialized in hosting NVIDIA's GPUs for bandwidth and AI throughput reasons
12:20PM EDT - "Amazing increase in system and memory bandwidth"
12:20PM EDT - And we're now deconstructing Jensen's kitchen...
12:20PM EDT - Grace in the artist-envisioned flesh
12:21PM EDT - NVIDIA has already lined up a customer for Grace: CSCS, who is building their Alps supercomputer
12:21PM EDT - Set to come online in 2023
12:21PM EDT - NVIDIA is now a CPU, GPU, and DPU company
12:22PM EDT - Each chip architecture will have a 2 year rhythm, with likely a kicker in-between
12:22PM EDT - NVIDIA will not stop supporting x86
12:22PM EDT - Instead they'll support both Arm and x86
12:24PM EDT - Speaking of Arm, NVIDIA is developing an Arm SDK, in a partnership with Ampere (the company)
12:24PM EDT - And jumping subjects again, this time to edge AI
12:27PM EDT - Recapping NVIDIA's various AI libraries and toolkits
12:27PM EDT - Which NVIDIA simply calls "NVIDIA AI"
12:27PM EDT - From PCs and laptops to workstations and supercomputers
12:28PM EDT - But one segment of the market that NVIDIA has not focused on up until now has been enterprise computing
12:28PM EDT - So NVIDIA is announcing their EGX enterprise platform
12:28PM EDT - NVIDIA AI runs on VMware
12:29PM EDT - So NVIDIA AI is available within virtualized environments
12:29PM EDT - "The missing link is 5G"
12:30PM EDT - NVIDIA is putting together another new hardware platform, which they are calling the Aerial A100
12:30PM EDT - An A100 GPU and Bluefield 2 processor on a single PCIe card
12:30PM EDT - For use in 5G basestations
12:30PM EDT - Software defined, with acceleration of PHY, crypto, packet processing, and more
12:31PM EDT - Which will be offered as part of an EGX edge server package
12:32PM EDT - Announcing NVIDIA Morpheus: a data center security product
12:32PM EDT - This is another DPU-centric product
12:32PM EDT - Now rolling an informational video about how NVIDIA is using Morpheus in-house
12:33PM EDT - Morpheus flags when it encounters unencrypted data
12:33PM EDT - Relying on AI, rather than specific pattern matching
12:35PM EDT - And recapping NVIDIA's enterprise hardware offerings, backed by EGX servers
12:36PM EDT - Now on to graphics-related AI projects like DLSS and variouos GANs
12:37PM EDT - NVIDIA sees the next wave of AI including increasingly plug-and-play use of the technology
12:37PM EDT - To that end, NVIDIA is adding even more pre-trained models to their collection for customers
12:38PM EDT - Announcing NVIDIA Tao framework
12:38PM EDT - And NVIDIA fleet commmand for securely controlling AI edge servers
12:39PM EDT - Now rolling a video about a customer using NVIDIA's Tao and Fleet Command products
12:40PM EDT - Starting with a pre-trained model, and then using Tao to re-train the model to better accomodate the specific job site
12:40PM EDT - All of the models are trained in minutes
12:40PM EDT - And the updated models are deployed via Fleet Command
12:41PM EDT - Pick a pre-trained model from NGC, optimize it with Tao, and then deploy it via Fleet Command
12:41PM EDT - Now on to conversational AIs
12:42PM EDT - NVIDIA's Jarvis package is now available for production use
12:42PM EDT - Jarvis has 90% recognition accuracy out of the box
12:42PM EDT - 5 languages supported today
12:43PM EDT - "No more mechanical talk"
12:43PM EDT - Jensen is focusing on the edge use cases for Jarvis, and where it could be run
12:44PM EDT - And NVIDIA is partnering with Mozilla to collect voice samples to better train Jarvis and other future voice AI systems
12:44PM EDT - "I have no idea what I said, but Jarvis recognized it perfectly"
12:45PM EDT - And showing Jarvis doing English to Japanese translations (voice to text to text)
12:45PM EDT - And configurable voice options, including intensity and enthusiasm
12:46PM EDT - Now on to recommender systems
12:46PM EDT - (We're moving at a breakneck pace here. NVIDIA has a lot of subjects to get through)
12:46PM EDT - Announcing NVIDIA Merlin, NVIDIA's end-to-end accelerated recommender system
12:47PM EDT - (A recommender system is exactly what it sounds like: a system that attempts to figure out what a user would prefer, and thus what they should be recommended)
12:47PM EDT - And on to NVIDIA Maxine, NVIDIA's video conferencing technology suites
12:48PM EDT - Which incorporates Jarvis voice recognition and translation
12:48PM EDT - Also showing off an eye contact faker/correcter
12:49PM EDT - A lot of people are videoconferencing these days, to say the least. So NVIDIA is keen on lining up customers in that market with tools to improve the experience
12:49PM EDT - Announcing NVIDIA Triton inference server
12:50PM EDT - Triton schedules models on to hardware. Any model and framework on to the appropriate hardware
12:51PM EDT - And a quick look at biomedical molecule simulations using Triton
12:52PM EDT - And now on to talking about what customers have been doing with NVIDIA's AI technologies
12:52PM EDT - Best Buy, Spotify, T-Mobile, and more
12:53PM EDT - And now on to automotive and DRIVE AV
12:53PM EDT - "AV computing demand is skyrocketing"
12:54PM EDT - Automakers still need more computing power
12:54PM EDT - Recapping NVIDIA's Orin SoC, which is set to arrive next year
12:56PM EDT - And the possibility of using a single Orin system as a central computer for everything within a car. From autonomous driving to dashes and infotainment, all execution segregated
12:56PM EDT - And NVIDIA's next-generation SoC past Orin is already in development
12:56PM EDT - DRIVE Atlan
12:56PM EDT - 1000 TOPS on a single chip
12:57PM EDT - Newly incorporating NVIDIA's AI and DPU technologies on top of the many other existing hardware features
12:57PM EDT - Due in 2025
12:58PM EDT - Now talking about NVIDIA's increasing number of major automotive customers, and what they're doing with NV's tech
12:58PM EDT - The big one, of course: robo taxis
12:59PM EDT - Driverless trucks, anyone?
12:59PM EDT - And now we're reaching the end, and Jensen is looping back to Omniverse
12:59PM EDT - Running NVIDIA DRIVE simulations within Omniverse
12:59PM EDT - And digital twin opportunities
01:00PM EDT - NVIDIA's Drive Sim engine will be available to Omniverse users
01:00PM EDT - Now rolling a video
01:01PM EDT - Showing Drive Sim in action, inside and outside of a simulated car
01:02PM EDT - And now on to the recap
01:03PM EDT - Omniverse
01:03PM EDT - DGX systems and Grace CPUs
01:03PM EDT - Jarvis, Merlin, and edge AI
01:04PM EDT - NVIDIA Tao, Fleet Command, and Triton
01:04PM EDT - And Drive, Orin, and the new Atlan SoC
01:05PM EDT - And that's a wrap. Thanks again for joining us
Post Your CommentPlease log in or sign up to comment.
View All Comments
SarahKerrigan - Monday, April 12, 2021 - linkIn which... Nvidia announces a Neoverse-V1 server CPU with integrated NVlink.
(Probably not, but it would be cool.)
FreckledTrout - Monday, April 12, 2021 - linkWhat about new spatulas?
SarahKerrigan - Monday, April 12, 2021 - linkI WAS KIDDING.
AdrianBc - Monday, April 12, 2021 - linkEither you had inside information or you made a very good guess prior to the presentation, because Nvidia has announced precisely that.
The only possible difference is that Nvidia did not specify whether the ARM core is Neoverse V1 or Neoverse N2, as both are successors to the current Neoverse cores.
SarahKerrigan - Monday, April 12, 2021 - linkNvidia's been hiring server CPU folks for a while and Power10 has no NVlink. Taken together, that pointed pretty strongly to a server CPU being announced at some point, though I didn't seriously think it would be this early (figured it would wait until after there's some resolution to the ARM acquisition.)
No inside information.
Operandi - Monday, April 12, 2021 - linkMy bet is egg whisks this time around. At least 25 of them, and a multitude of colors of course.
WaltC - Monday, April 12, 2021 - linkFrom the keynote:
"I am so happy today to announce nVidia's new business model: Vaporware GPUs! From now on we can paper launch any GPU with any specifications we care to create without the added burden of having to actually make and ship the hardware that will do all the wonderful things we claim it will do! And people will believe it--because we can blame the miners, the FABs, and the UFOs for the almost total shortages! It Just Works! I'm so happy today! Thank you very much. Our next vaporware GPUs coming in 3-2-1..."
Unashamed_unoriginal_username_x86 - Monday, April 12, 2021 - linkYeah the profit margin from 0 expenses on 0 revenue is insane. Expecting more companies to pretend to make products soon
mrvco - Monday, April 12, 2021 - linkSorry, but I'm tuned out until general availability of gaming GPUs improves to a point where either lottery levels of luck and/or enriching scalpers is no longer mandatory.
Unashamed_unoriginal_username_x86 - Monday, April 12, 2021 - linkGood, this keynote isn't for gaming GPUs.