All News | Boards | Chips | Devices | Software | LinuxDevices.com Archive | About | Contact | Subscribe
Follow LinuxGizmos:
Twitter Facebook Pinterest RSS feed
*   get email updates   *

Nvidia unveils high-end Jetson AGX Orin plus an Omniverse platform for AR/VR

Nov 9, 2021 — by Eric Brown 497 views

[Updated: Nov. 15] — Nvidia unveiled a “Jetson AGX Orin” module with 12x 2GHz Cortex-A78 and 2,048 Ampere GPU cores for up to 200-TOPS AI plus 32GB LPDDR5 and 64GB eMMC. The AGX Orin runs some of Nvida’s new “Omniverse” applications for AR, VR, and robotics.

At its GTC 2021 Fall conference, Nvidia announced its most advanced Jetson module yet, delivering 6x time the processing power of the Jetson AGX Xavier while maintaining pin compatibility and the 100 x 87mm form factor. Due in 1Q 2022, the Jetson AGX Orin features 12x, 2GHz Cortex-A78 cores along with 2,048 1GHz Ampere GPU cores. The module also provides 64 Tensor cores, dual NVDLA 2.0 deep learning accelerators, and a PVA v2.0 vision accelerator to help deliver up to 200-TOPS AI performance using INT8.



Jetson AGX Orin with dev kit (left) and Nvidia’s Omniverse Avatar demonstrated with a Toy Jensen Omniverse Avatar featuring Nvidia CEO Jensen Huang. (If he’s looking for Mark Zuckerberg, he’s in the wrong metaverse.)
(click images to enlarge)

The Jetson AGX Orin achieves the 6x power boost within a 15W to 50W consumption range, compared to 30W for the AGX Xavier. (Presumably, the benchmark was running at 50W.) The upper RAM limit remains at 32GB, and eMMC storage capacity have doubled to 64GB.

The Jetson AGX Orin is designed for advanced robotics and AI edge applications for manufacturing, logistics, retail, service, agriculture, smart city, healthcare, and life science. Software support remains the same, including Ubuntu Linux, Jetpack, CUDA-X, and most of the other Nvidia AI platforms such as the recently updated TensorRT 8.0 for Nvidia GPUs.

— ADVERTISEMENT —


As usual at GTC, the Jetson announcement was a side dish to more server/cloud oriented Nvidia software news, in this case the official launch of Nvidia’s Omniverse platform for augmented reality and virtual reality. The AGX Xavier is powerful enough to run many of the Omniverse applications (see farther below).



Block diagram for AGX Orin’s Orin SoC (left) and for AGX Orin module
(click images to enlarge)

The Jetson AGX Xavier is built around an Orin SoC, which is already used by Nvidia’s Drive AGX Orin autonomous vehicle computer. The Drive AGX Orin was announced in Dec. 2019, just as the AGX Xavier SoC that drives the Jetson AGX Xavier was first announced in 2017 in Nvidia’s Drive PX Pegasus computer for Level 5 self-driving cars.

The Orin SoC’s 12x Cortex-A78AE cores are split into 3x quad-core blocks, each with 2MB cache. The SoC also offers 6MB L3. The SoC use the same safety-critical Cortex-A78AE variant used in the Drive AGX Orin, with the AE referring to “Automotive Enhanced.” Cortex-A78AE offers a “Split Mode” for ensuring ASIL-B functional safety compliance, as well as a traditional, lockstep “Locked Mode” and a “Hybrid Mode” in between.

As usual, the chief attraction on the Jetson is the GPU, in this case the 2,048-core Ampere GPU, which is used on high-end Nvidia graphics cards such as the T4 and A100. Ampere features a fine-grained compute structure called Sparsity that is claimed to double throughput and reduce memory usage. The Ampere architecture integrates dual Graphic Processing Clusters (GPCs), 8x Texture Processing Clusters (TPCs), and 16x Streaming Multiprocessors (SMs). It also offers 192KB of L1 cache per SM and 4MB of L2 cache.



Block diagram for Orin SoC’s Ampere GPU (left) and more detailed AGX Orin diagram
(click images to enlarge)

By comparison, the earlier AGX Xavier has 8x 2.26GHz ARMv8.2 CPU cores, a 512-core, 1.37GHz Nvidia Volta GPU, 64 tensor cores, and a 7-way VLIW vision chip. The AGX Xavier can achieve 32-TOPS AI performance. The 6x performance boost, which is probably based on benchmarks running at 50W, is achieved despite the reduction in clock rate from the AGX Xavier.

The up to 32GB LPDDR5 offers faster 205 GB/sec throughput compared to the Xavier’s 137GB/sec. There is now 64GB eMMC 5.1, up from 32GB eMMC 5.0 on the Xavier.

The Jetson AGX Orin supports up to 6x cameras via 16-lanes of MIPI-CSI-2 and offers video encode up to 2x 4Kp60. Video decode is listed at up to 8Kp60 via multi-mode DP 1.4a. The module also supports eDP 1.4a and HDMI 2.1.



Jetson AGX Orin (left) and I/O specs
(click images to enlarge)

The AGX Orin provides controllers for 4x 10GbE and a single GbE port and offers PCIe Gen4 root and endpoint interfaces for 2x PCie x8, 1x x4, and 2x x1. Other peripheral support includes 3x USB 3.2 (see chart above).

A Jetson AGX Orin Developer Kit will be available with a PCIe x16 slot that supports PCIe Gen4 x8. The other x8 is used to drive 3x 10GbE ports, which joins a standard RJ45 port that can be used for a fourth 10GbE.



Jetson AGX Orin Developer Kit and major specs
(click images to enlarge)

The 110 x 110 x 71.65mm dev kit is further equipped with a microSD slot and M.2 M-key for NVMe storage. An M.2 E-key is preloaded with WiFi/BT, which is not available on the AGX Orin module itself. Other features include MIPI-CSI-2, DP++, 5x USB, a 40-pin GPIO header, and more, as shown in the chart above.

 
Nvidia’s Omniverse

Nvidia’s main course at GTC 2021 Fall is a 3D development platform called Omniverse, which is used to develop augmented reality and virtual reality applications. Nvidia launched a beta of Omniverse last December, which has been downloaded by more than 70,000 developers.

The company has now officially launched the platform along with some specific Omniverse Avatar and Omniverse Replicator applications and new features across the Omniverse product line such as and multi-GPU and AR/VR enhancements. There are also Omniverse integrations for infrastructure and industrial digital-twin applications for the engineering and construction industry, via software from Bentley Systems and Esri.



Omniverse Replicator images for Isaac SIM robotics (left) and Nvidia Drive
(click image to enlarge)

Many of the Omniverse offerings are available on the AGX Orin. These include an “Nvidia Isaac Sim on Omniverse” robotics development platform with enhanced ROS 2 support, as well as its related Omniverse Replicator software for synthetic data generation. The Omniverse Replicator for Isaac SIM is defined as a cockpit for ML engineers with production-quality datasets at scale that match target distributions.

At a press pre-briefing, Deepu Talla, VP and GM of Nvidia’s Embedded and Edge Computing unit, promoted the robotics platform for offering labeled data, domain randomizaton, and the ability to change lighting and camera angles to make simulated models. Talla said that the Omniverse platform in general would be sold to enterprises as private platforms starting at $9K per year.

The Jetson AGX Orin does not appear to support the new Omniverse Avatar platform for building interactive 3D avatars. Omniverse Avatar is being showcased in in a Project Maxine for Omniverse Avatar that connects computer vision, Riva speech AI, and avatar animation and graphics into a real-time conversational AI robot. At GTC, Nvidia CEO Jensen Huang demonstrated the Toy Jensen Omniverse Avatar, as shown in the image of at top. There is also a related Project Tokkio application that could, for example, present customers with an interactive, animated AI that answers questions about a restaurant menu.

The AGX Xavier supports the latest version of Nvidia Drive for autonomous driving, as well as new Drive Concierge and Drive Chauffeur AI platforms for AI assistants for passenger infotainment. Jetson AGX Orin also supports a new Nvidia Clara Holoscan SDK for healthcare applications. These do not appear to be officially part of the Omniverse platform, although there is an Omniverse for Drive SIM Replicator that that is much like the Isaac Replicator, but for autonomous vehicles.

Although Omniverse supports gaming and entertainment applications, it can be considered a more enterprise and industrial counterpoint to Facebook/Meta’s plans for developing Metaverse software. Since Nvidia is not faced with the high level of criticism and government investigations Facebook has encountered about questionable business practices, spreading misinformation, and in general making us all miserable, the company has declined to change its company name to Omni.

 
Further information

The Jetson AGX Orin and its dev kit will be available in Q1 2022, with pricing undisclosed. More information may be found in Nvidia’s AGX Orin announcement and product page.

More on Nvidia Omniverse may be found in this Omniverse announcement focusing on Omniverse Replicator, as well as this product page. More on other Omniverse offerings and other GTC announcements may be found in Nvidia’s GTC 2022 Fall overview announcement.
 

(advertise here)


Print Friendly, PDF & Email
PLEASE COMMENT BELOW

Please comment here...