NVIDIA: summary of the 30 announcements made at GTC 2021 – Tom’s Hardware

Automotive, supercomputing, and digital collaboration announcements; but the highlight of the show is Toy Jensen.

NVIDIA CEO Jensen Huang gave a keynote address at GTC (GPU Technology Conference) rich in announcements: to report them, NVIDIA shares 6 press releases and 27 blog posts; which is to say that it is quite difficult to navigate, especially since they cover a large number of sectors, from AI avatars to autonomous vehicles, including forest fire forecasting tools (automotive, supercomputing, and digital collaboration to be precise). You can access a summary in English here and watch the full talk or a video summarizing the main announcements below.

For our part, we will focus mainly on the announcements relating to the hardware and then evoke other subjects, without however claiming to be exhaustive. The descriptions of NVIDIA come from various press releases in French.

NVIDIA Quantum-2, the next generation of InfiniBand network platform

New Generation NVIDIA Quantum-2 InfiniBand Network Platform Brings Greater Accessibility and Enhanced Security to Service Providers cloud computing and supercomputing centers. It consists of the NVIDIA Quantum-2 switch, ConnectX-7 network adapter, BlueField-3 data processing unit (DPU), and all software that supports the new architecture.

We had already mentioned the DPU BlueField-3 last spring. Regarding the switch Quantum-2, it achieves a speed of 400 Gbit / s. The company argues that it “Doubles the network speed and triples the number of network ports.” It accelerates performance 3 times and reduces the need for data center switches by 6 times, while reducing power consumption and data center space by 7% each ”.

This switch engraved in 7 nm has 57 billion transistors, a little more than the NVIDIA A100 GPU (54 billion transistors). It features 64 ports at 400 Gbps or 128 ports at 200 Gbps and will be offered in a variety of switching systems up to 2,048 ports at 400 Gbps or 4096 ports at 200 Gbps, more than five times the switching capacity. from the previous generation, Quantum-1.

NVIDIA DOCA 1.2

NVIDIA announced the creation of its “zero-trust” framework, which “Enables developers to deliver cybersecurity services on an ‘as-a-service’ basis, extending data center threat protection to nearly every touchpoint between applications, data, devices, storage, infrastructure and users’.

In practice, “NVIDIA DOCA 1.2 Enables Developers to Build Applications for NVIDIA BlueField DPUs […]. With the DOCA 1.2 zero-trust security framework, developers can create framed cloud services that control access to resources, validate every application and user, isolate potentially compromised machines, and help protect data from breaches and theft ”.

Jeston AGX Orin

NVIDIA also unveiled its new robotic computer Jetson AGX Orin. With dimensions of 100mm x 87mm, Jetson AGX Orin offers up to 6 times the performance of Jetson AGX Xavier in the same form factor. Its AI performance is 200 TOPS (INT8). Built on the NVIDIA Ampere architecture, it is “The world’s smallest, most powerful and energy efficient AI supercomputer for robotics, autonomous machines, medical devices and other forms of on-board computing at the edge.” according to NVIDIA.

In addition, NVIDIA announced a synthetic data generation engine, Omniverse Replicator, which generates synthetic training data for robots. The firm is supporting it with a new Omniverse Replicator application for Isaac sim. “New Capabilities in Isaac Sim Allow ML Engineers to Build Production-Quality Synthetic Data Sets to Train Robust Deep Learning Perception Models”.

Besides, “The latest version of Isaac also includes significant support for the Robot Operating System (ROS) developer community. NVIDIA has released six new Isaac GEM, hardware-accelerated software packages that make it easier for ROS developers to build high-performance AI-based robots on the Jetson platform ”.

NVIDIA Omniverse

In another register, the company is expanding the capabilities of NVIDIA Omniverse and launches Omniverse Enterprise. The platform now offers rendering functions in AR, VR and Multi-GPU.

“The realm of virtual worlds is taking a big leap forward today with the introduction of new features and partners for NVIDIA Omniverse – accelerating the Internet’s evolution from ordinary 2D to a dazzling world that comes to life in 3D . Omniverse – a platform that serves as the connective tissue for physically precise 3D virtual worlds – is enriched with new features such as AR, VR and multi-GPU rendering, as well as integrations for digital twinning applications. infrastructure and industrial with software from Bentley Systems and Esri. “

More generally, Omniverse enables engineers and designers to build physically accurate digital twins of buildings and products, or create massive and realistic simulation environments to train robots or autonomous vehicles before they are deployed in the physical world. . Since launching its open beta in December, Omniverse has been downloaded by over 70,000 individual creators. It is also used by professionals at over 700 companies, including BMW Group, CannonDesign, Epigraph, Ericsson, architectural firms HKS and KPF, Lockheed Martin and Sony Pictures Animation ”.

Below, some concrete applications.

https://www.youtube.com/watch?v=a5w_aZfncO0

Omniverse Avatar, a platform for creating AI avatars

To support the enthusiasm – whether or not it is real, the future will tell – around the metaverse, NVIDIA has also demonstrated its solution. Omniverse Avatar, which is part of NVIDIA Omniverse. Basically, Omniverse Avatar allows you to model and animate AI avatars.

An explanation a little more elaborated by NVIDIA: “Omniverse Avatar connects the company’s technologies in voice AI, computer vision, natural language understanding, recommendation engines and simulation technologies. The avatars created in the platform are interactive characters with 3D ray-traced graphics, able to see, speak, converse on a wide range of subjects and understand naturally expressed intentions. “

Concretely, we were able to discover Toy Jensen, avatar of Jensen Huang in person, answer a few questions; attend a demonstration of the Tokkio project, able to take orders from customers in a restaurant; discover Poject Maxine, who adds “From video and audio functions to virtual collaboration and content creation applications”.

Fighting forest fires with AI

Finally, since we touched on it in the first paragraph, here are a few words about how NVIDIA Omniverse can help fight wildfires.

To better understand forest fires and stop their spread, NVIDIA and Lockheed Martin today announced their collaboration with the Forest Services of the United States Department of Agriculture and the Colorado Division of Fire Prevention and Control (DFPC) using AI and digital simulation. The companies are also announcing the construction of the world’s first AI lab dedicated to forest fire prediction and response. This lab will use NVIDIA’s AI infrastructure and NVIDIA Omniverse advanced visualization and digital twin simulation platform to process the scale of a fire and predict its progression. By recreating the fire via a physically accurate digital twin, the system will be able to suggest actions to better extinguish the fire. “