10:58AM EDT – Welcome GPU watchers to another GTC spring keynote

10:59AM EDT – We’re here reporting from the comfort of our own homes on the latest in NVIDIA news from the most important of the company’s GTC keynotes

10:59AM EDT – GTC Spring, which used to just be *the* GTC, is NVIDIA’s traditional venue for big announcements of all sorts, and this year should be no exception

11:00AM EDT – NVIDIA’s Ampere server GPU architecture is two years old and arguably in need of an update. Meanwhile the company also dropped a bomb last year with the announcement of the Grace CPU, which would be great to find out more about

11:00AM EDT – And here we go

11:01AM EDT – For the first time in quite a while, NVIDIA is not kicking things off with a variant of their « I am AI » videos

11:02AM EDT – Instead, we’re getting what I’m guessing is NV’s « digitial twin » of their HQ

11:02AM EDT – And here’s Jensen

11:03AM EDT – And I spoke too soon. Here’s a new « I am AI » video

11:04AM EDT – These videos are arguably almost cheating these days. In previous years NV has revealed that they have AI doing the music composition and, of course, the voiceover

11:06AM EDT – And back to Jensen. Discussing how AI is being used in biology and medical fields

11:07AM EDT – None of this was possible a decade ago

11:07AM EDT – AI has fundamentally changed what software can make. And how you make software

11:07AM EDT – The next wave of AI is robotics

11:08AM EDT – Digital robotics, avatars, and physical robotics

11:08AM EDT – And omniverse will be essential to making robotics

11:08AM EDT – In the past decade, NV’s hardware and software have delivered a million-x speedup in AI

11:09AM EDT – Today NVIDIA accelerates millions of developers. GTC is for all of you

11:10AM EDT – Now running down a list of the major companies who are giving talks at GTC 2022

11:10AM EDT – Recapping NVIDIA’s « Earth 2 » digital twin of the Earth

11:12AM EDT – And pivoting to a deep learning weather model called FourCastNet

11:12AM EDT – Trained on 4TB of data

11:13AM EDT – FourCastNet can predict atmospheric river events well in advance

11:13AM EDT – Now talking about how NVIDIA offers multiple layers of products, across hardware, software, libraries, and more

11:14AM EDT – And NVIDIA will have new products to talk about at every layer today

11:14AM EDT – Starting things off with a discussion about transformers (the deep learning model)

11:14AM EDT – Transformers are the model of choice for natural language processing

11:15AM EDT – e.g. AI consuming and generating text

11:15AM EDT – (GPT-3 is scary good at times)

11:15AM EDT – « AI is racing in every direction »

11:16AM EDT – Rolling a short video showing off a character model trained with reinforcement learning

11:16AM EDT – 10 years of simulation in 3 days of real world time

11:17AM EDT – NV’s hope is to make animating a character as easy as talking to a human actor

11:18AM EDT – Now talking about the company’s various NVIDIA AI libraries

11:18AM EDT – And Triton, NVIDIA’s inference server

11:19AM EDT – Triton has been downloaded over a million times

11:20AM EDT – Meanwhile, the Riva SDK for speech AI is now up to version 2.0, and is being released under general availability

11:20AM EDT – « AI will reinvent video conferencing »

11:21AM EDT – Tying that in to Maxine, NV’s library for AI video conferencing

11:22AM EDT – Rolling a short demo video of Maxine in action

11:22AM EDT – It sounds like Maxine will be a big item at the next GTC, and today is just a teaser

11:23AM EDT – Now on to recommendation engines

11:23AM EDT – And of course, NVIDIA has a framework for that: Merlin

11:24AM EDT – The Merlin 1.0 release is now ready for general availability

11:24AM EDT – And back to transformers

11:24AM EDT – Google is working on Switch, a 1.6 trillion parameter transformer

11:25AM EDT – And for that, NVIDIA has the Nemo Megatron framework

11:26AM EDT – And now back to where things started, AI biology and medicine

11:26AM EDT – « The conditions are prime for the digital biology revolution »

11:28AM EDT – (This black background does not pair especially well with YouTube’s overly compressed video streams)

11:28AM EDT – And now on to hardware

11:28AM EDT – Introducing NVIDIA H100!

11:28AM EDT – 80B transistor chip built on TSMC 4N

11:28AM EDT – 4.9TB/sec bandwidth

11:28AM EDT – First PCIe 5.0 GPU

11:28AM EDT – First HBM3 GPU

11:28AM EDT – A single H100 sustains 40TBit/sec of I/O bandwidth

11:29AM EDT – 20 H100s can sustain the equivalent of the world’s Internet traffic

11:29AM EDT – Hopper architecture

11:29AM EDT – « 5 groundbreaking inventions »

11:29AM EDT – H100 has 4 PFLOPS of FP8 perform

11:29AM EDT – 2 PFLOPS of FP16, and 60 TFLOPS of FP64/FP32

11:30AM EDT – Hopper’s FP8 is 6x the performance of Ampere’s FP16 perf

11:30AM EDT – Hopper introduces a transformer engine

11:30AM EDT – Transformer Engine: a new tensor core for transformer training and inference

11:31AM EDT – On the security front, Hopper adds full isolation for MIG mode

11:31AM EDT – And each of the 7 instances is the performance of two T4 server GPUs

11:31AM EDT – The isolated MIG instances are fully secured and encrypted. Confidential computing

11:32AM EDT – Data and application are protected during use on the GPU

11:32AM EDT – Protects confidentiality of valuable AI models on shared or remote infrastructure

11:32AM EDT – New set of instructions: DPX

11:32AM EDT – Designed to accelerate dynamic programming algorithms

11:33AM EDT – Used in things like shortest route optimization

11:33AM EDT – Hopper DPX instructions will speed these up upwards of 40x

11:33AM EDT – COWOS 2.5 packaging

11:33AM EDT – HBM3 memory

11:34AM EDT – 8 SXMs are paired with 4 NVSwitch chips on an H100 HGX board

11:34AM EDT – Dual « Gen 5 » CPUs

11:34AM EDT – Networking provided by Connectx-7 NICs

11:35AM EDT – Introducing the DGX H100, NVIDIA’s latest AI computing system

11:35AM EDT – 8 H100 GPUs in one server

11:35AM EDT – 640GB of HBM3

11:35AM EDT – « We have a brand-new way to scale up DGX »

11:35AM EDT – NVIDIA NVLink Switch System

11:36AM EDT – Connect up to 32 nodes (256 GPUs) of H100s via NVLink

11:36AM EDT – This is the first time NVLink has been available on an external basis

11:36AM EDT – Connects to the switch via a quad port optical transceiver

11:36AM EDT – 32 transceivers connect to a single node

11:37AM EDT – 1 EFLOPS of AI performnace in an 32 node cluster

11:37AM EDT – And DGX SuperPods scale this up further with the addition of Quantum-2 Infiniband

11:37AM EDT – NVIDIA is building another supercomputer: Eos

11:38AM EDT – 18 DGX Pods. 9 EFLOPS FP16

11:38AM EDT – Expect Eos to be the fastest AI computer in the world, and the blueprint for advanced AI infrastrucutre for NVIDIA’s hardware partners

11:38AM EDT – Standing up Eos now and online in a few months

11:39AM EDT – Now talking about performance

11:39AM EDT – 6.3x transformer training performance

11:39AM EDT – And 9x on a different transformer

11:39AM EDT – H100: the new engine of the world’s AI infrastructure

11:39AM EDT – « Hopper will be a game changer for mainstream systems as well »

11:40AM EDT – Moving data to keep GPUs fed is a challenge

11:40AM EDT – Attach the network directly to the GPU

11:40AM EDT – Announcing the H100 CNX

11:40AM EDT – H100 and a CX-7 NIC on a single card

11:41AM EDT – Skip the bandwidth bottlenecks by having the GPU go directly to the network

11:41AM EDT – Now on to Grace

11:41AM EDT – Grace is « progressing fantastically » and is on track to ship next year

11:41AM EDT – Announcing Grace Hopper, a single MCM with a Grace CPU and a Hopper GPU

11:42AM EDT – The chips are using NVLink to communicate

11:42AM EDT – Announcing Grace CPU Superchip. Two Graces in MCM

11:42AM EDT – 144 CPU cores, 1TB/sec of LPDDR5X

11:42AM EDT – Connected via NVLink Chip-2-Chip (C2C)

11:43AM EDT – Entire module, including memory, is only 500W

11:43AM EDT – And all of NVIDIA’s software platforms will work on Grace

11:43AM EDT – Now talking about the NVLink C2C link used to connect these chips

11:44AM EDT – NVLink C2C allows for many different Grace/Hopper configurations

11:45AM EDT – And NVIDIA is opening up NVLink to let customers implement it as well to connect to NVIDIA’s chips on a single package

11:45AM EDT – So NV is going chiplet and semi-custom?

11:45AM EDT – Now on to NVIDIA SDKs

11:47AM EDT – Over 175 companies are testing NVIDIA CuOpt

11:47AM EDT – NVIDIA DGL Container: training large graph NNs across multiple nodes

11:48AM EDT – NVIDIA cuQuantum: SDK for accelerating quantum circuits

11:48AM EDT – Aerial: SDK for 5G radio

11:49AM EDT – And NV is already getting ready for 6G

11:49AM EDT – Sionna: new framework for 6G research

11:50AM EDT – Monai: AI framework for medical imaging

11:50AM EDT – Flare: SDK for federated learning

11:51AM EDT – « The same NVIDIA systems you already own just got faster »

11:51AM EDT – And that’s the word on NVIDIA’s massive (and growing) library of frameworks

11:52AM EDT – Now talking about the Apollo 13 disaster, and how the fully functional replica on Earth helped to diagnose and deal with the issue

11:52AM EDT – Thus coining the term « digital twin »

11:53AM EDT – « Simulating the world is the ultimate grand challenge »

11:53AM EDT – Dovetailing into omniverse

11:53AM EDT – And what Omniverse is useful for today

11:53AM EDT – Industiral digital twins and more

11:54AM EDT – But first, the technologies that make omniverse possible

11:54AM EDT – Rendering, materials, particle simulations, physics simulations, and more

11:55AM EDT – « Omniverse Ramen Shop »

11:55AM EDT – (Can’t stop for lunch now, Jensen, there’s a keynote to deliver!)

11:56AM EDT – Omniverse is scalable from RTX PCs to large systems

11:56AM EDT – « But industrial twins need a new type of purpose-built computer »

11:56AM EDT – « We need to create a synchronous datacenter »

11:56AM EDT – NVIDIA OVX Server

11:57AM EDT – 8 A40s and dual Ice Lake CPUs

11:57AM EDT – And NVIDIA OVX SuperPod

11:57AM EDT – Nodes are synchronized

11:57AM EDT – Announcing Specturm-4 Switch

11:57AM EDT – 400G Ethernet Switch

11:58AM EDT – 100B transistors(!)

11:58AM EDT – World’s first 400G end-to-end networking platform

11:58AM EDT – Timing precision to a new nanoseconds

11:59AM EDT – The backbone of their omniverse computer

11:59AM EDT – Samples in late Q4

12:00PM EDT – NVIDIA is releasing a major Omniverse kit at this GTC

12:00PM EDT – Omniverse Avatar: a framework for building avatars

12:01PM EDT – Showing off Toy Jensen

12:01PM EDT – According to Jensen, TJ is not pre-recorded

12:01PM EDT – NV has also replicated a version of Jensen’s voice

12:02PM EDT – The tex to speech is a bit rough, but it’s a good proof of concept

12:03PM EDT – (At the rate this tech is progressing, I half-expect Jensen to replace himself with an avatar in GTC keynotes later this decade)

12:04PM EDT – Discussing all of the technologies/libraries that went into making Toy Jensen

12:04PM EDT – The next wave of AI is robotics

12:04PM EDT – Drive is being lumped in here, along with Isaac

12:06PM EDT – Now rolling a demo video of Drive with an AI avatar

12:08PM EDT – Showing the various sub-features of Drive in action. As well as how the system monitors the driver

12:08PM EDT – And parking assistance, of course

12:09PM EDT – Hyperion is the architecture of NVIDIA’s self-driving car platform

12:09PM EDT – Hyperion 8 can achieve full self driving with the help of its large suite of sensors

12:09PM EDT – And Hyperion 9 is being announced today for cars shipping in 2026

12:10PM EDT – H9 will process twice as much sensor data as H8

12:10PM EDT – NVIDIA Drive Map

12:10PM EDT – NV expects to map all major highways in North America, Europe, and Asia by the end of 2024

12:12PM EDT – And all of this data can be loaded into Omniverse to create a simulation environment for testing and training

12:12PM EDT – Pre-recorded Drive videos can also be ingested and reconstructed

12:13PM EDT – And that’s Drive Map and Drive Sim

12:13PM EDT – Now on to the subject of electric vehicles

12:13PM EDT – And the Orin SoC

12:14PM EDT – Orin started shipping this month (at last!)

12:14PM EDT – BYD, the second-largest EV maker globally, will adopt Orin for cars shipping in 2023

12:14PM EDT – Now on to NV medical projects

12:15PM EDT – Using NVIDIA’s Clara framework to process copius amounts of microscope data in real time

12:16PM EDT – Clara Holoscan

12:17PM EDT – Announcing Clara Holoscan MGX platform, medical grade readiness in Q1 2023

12:18PM EDT – NVIDIA Metropolis and Isaac

12:18PM EDT – « Metropolis has been a phenomenal success »

12:19PM EDT – Customers can use Omniverse to make digital twins of their facilities

12:19PM EDT – PepsiCo has built digital twins of their packaging and distribution centers

12:20PM EDT – Major release of Isaac: Isaac for AMRs

12:21PM EDT – Isaac Nova: reference AMR robot system

12:21PM EDT – Announcing Jetson Orin developer kits

12:21PM EDT – Nova AMR available in Q2

12:23PM EDT – And using reinforcement learning to train robots in simulation, and then using that training data to program a real robot

12:23PM EDT – Train your robots in a sim until they’re ready to move into the real world

12:26PM EDT – And now a demo of how Amazon is using a digital twin

12:26PM EDT – Omniverse is helping Amazon optimize and simplify their processes

12:27PM EDT – Aggregating data from multiple CAD systems

12:27PM EDT – Test in the digital twin optimization concepts to see if and how it works

12:28PM EDT – And training models far faster in simulation than they could be trained in the real world in real time

12:29PM EDT – Announcing Omniverse Cloud

12:29PM EDT – « One click design collaboration »

12:29PM EDT – Rolling a demo

12:30PM EDT – Streaming Omniverse from GeForce Now, so Omniverse can be accessed on non-RTX systems

12:31PM EDT – And using Toy Jensen, having it modify an omniverse design via voice commands

12:32PM EDT – Omniverse: for the next wave of AI

12:32PM EDT – Jensen is now recapping today’s announcements

12:34PM EDT – H100 is in production with availability in Q3

12:35PM EDT – NVLink is coming to all future NV chips

12:35PM EDT – Omniverse will be integral for action-oriented AI

12:38PM EDT – « We will strive for another million-x in the next decade »

12:38PM EDT – And thanking NV employees, partners, and families

12:38PM EDT – And now for one more thing with Omniverse, which was used to generate all of the renders seen in this keynote

12:39PM EDT – A bit of musical fanfare to close things out

12:42PM EDT – And that’s a wrap. Please be sure to check out our Hopper H100 piece: https://www.anandtech.com/show/17327/nvidia-hopper-gpu-architecture-and-h100-accelerator-announced