Nvidia unveiled a scaled-down “Jetson Orin NX” sibling to its high-end AGX Orin module and officially launched its “Jetson AGX Orin Developer Kit.” The Orin NX is claimed to offer 3X the performance of the AGX Xavier. Back in November, Nvidia announced its latest flagship Jetson module, the Jetson AGX Orin, featuring 12x 2.2GHz Cortex-A78AE and 2,048 Ampere GPU cores. The company has now followed up with a dev kit for the module that is open for orders at $1,999 with a 32GB AGX Orin module, with shipments within 8x weeks. Most of the features of the dev kit were already announced in November, but there are a few new details (see farther below).

Nvidia also announced a reduced-feature, hexa- or octa-core Orin NX variant that is form factor compatible with the Jetson Xavier NX. Orin NX, which runs Ubuntu with Nvidia JetPack like all the other Jetson modules, is claimed to offer almost 5x the performance of the Xavier NX and 3x the performance of the AGX Xavier (see farther below). Jetson AGX Orin (left) and new Jetson Orin NX The products were announced at the GTC 2022 Spring conference where Nvidia also unveiled a 4nm Nvidia H100 GPU for datacenters that uses a new Hopper architecture, The H100 features PCIe Gen5 support, a Transformer Engine, and an NVLink interconnect for massive AI models and digital twins. There were also more details on the Grace CPU that Nvidia announced in November, and which together with the Hopper architecture celebrates computer pioneer Grace Hopper.

In addition to launching the AGX Orin dev kit, Nvidia teased an Isaac Nova Orin reference platform for developing 3D-sensing robots. The robotics kit is driven by a pair of AGX Orin modules and offers 2x back-depth and 2x front-depth 2-megapixel cameras plus 4x wide-angle surround cameras, also at 2MP. The kit also includes 2D Lidar at 25fps and surround ultrasonic sensors.

Isaac Nova Orin integrates with the Nvidia Omniverse AR/VR platform that was announced last November. Nvidia offered more details on Omniverse this week, including an Nvidia Modulus platform based on Omniverse, which is used to developer scientific digital twins.

Before we get into the Jetson Orin NX and dev kit, we will recap the Jetson AGX Orin, including a few new details supplied by Nvidia. The AGX Orin is claimed to deliver 8x time the processing power of the Jetson AGX Xavier while maintaining pin compatibility and the 100 x 87mm form factor.

The Jetson AGX Orin is built around Nvidia’s Orin SoC, which is already used by its Drive AGX Orin autonomous vehicle computer. In addition to the 12x 2GHz Cortex-A78AE and 2,048 Ampere GPU cores, the AGX Orin SoC provides 64 Tensor cores, dual NVDLA 2.0 deep learning accelerators, and a PVA v2.0 vision accelerator. There is also 3MB of L2 and 6MB of L3 cache. Block diagram for AGX Orin’s Orin SoC (left) and for AGX Orin module (click images to enlarge) In November, Nvidia said the upper RAM limit remained at the AGX Xavier’s 32GB. However, Nvidia now shows a new 64GB LPDDR4x model in addition to the 32GB RAM model. Both offer 64GB eMMC, twice that of the AGX Xavier.

As noted by Hackster.io, despite the AGX Orin’s greater processing power compared to the AGX Xavier, the new module is limited to half the video encode and decode support. Like the AGX Xavier, the module has 16-lanes of MIPI-CSI-2, but it is limited to 16 virtual CSI channels instead of 32. It lacks its predecessor’s 8x lanes of SLVS-EC (Scalable Low Voltage Signaling – Embedded Clock), a high-speed camera interface from Sony. Available encode formats include 2x 4K60, 4x 4K30, 8x 1080p60, and 16x 1080p30 (H.265).

Similarly, while the AGX Xavier supports 2x 8K40 video streams, the AGX Orin offers 1x 8K30, 3x 4K60, 6x 4K30, 12x 1080p60, and 24x 1080p30 (H.265) decoding options. On the other hand, you get 22 lanes of PCIe Gen4 vs. 16 lanes plus support for up to 4x 10GbE ports compared to the single GbE controller found on the AGX Xavier. (For more details on the AGX Orin, see our original Jetson AGX Orin report.)

  Up to 275 Sparse TOPS and 138 Dense TOPS

In November, Nvidia said the AGX Orin delivers up to 200-TOPS AI performance using INT8. The company has now split its TOPS specs into two parts — Sparse and Dense — based on the Ampere GPU’s ability to use a fine-grained compute structure called Sparsity that is claimed to double throughput and reduce memory usage. Nvidia now says the new, $1,599 (in 1K units) 64GB RAM model delivers 275 Sparse and 138 Dense INT8 TOPS ratings while the $899 32GB model has 200 Sparse/100 Dense TOPS.

According to an Nvidia blog post about sparsity, the technique compresses and removes “many zeros or values that will not significantly impact a calculation.” As a result, performance improves significantly while only marginally affecting accuracy.

AMD Xilinx published a recent blog post on Sparse vs. Dense TOPS that says that sparsity, or applying “zero-compression in sparse matrices,” can theoretically double system performance. However, the post also notes that sparsity methods “usually cannot be implemented in a practical system, so take any performance claims with a grain of salt.” Xilinx goes on to warn that you should never compare Sparse and Dense scores and suggests that TOPS, which is further complicated by power consumption, is overrated as a reliable measure of AI performance.

Xilinx is now listing sparse and dense TOPS ratings for its AI-enabled, but more FPGA-like Versal SoC platform, which can be considered a rival to the AGX Orin. The scores range from 11 Sparse/7 Dense TOPS to 342 Sparse/228 Dense TOPS, depending on the Versal model.

The Orin NX uses the same Cortex-A78AE CPU cores used by the 12x core AGX Xavier. The Xavier NX-sized module also has a 1024-core Ampere GPU with 32 tensor cores compared to 2,048 Ampere GPU cores and 64 tensor cores on the AGX Orin. We saw no details on cache, eMMC, CPU clock rate, or I/O support. Nvidia’s comparison chart between Orin NX and Xavier NX (click image to enlarge) The Orin NX comes in two models. The hexa-core Orin NX 8GB model integrates 8GB LPDDR4x and has a 10-20W TDP. It sells for $399 when bought with a 1K order. The octa-core Orin NX 16GB has 16GB RAM and 10-25W consumption and sells for $599.

The Orin NX 8GB has 70 Sparse/35 Dense TOPS while the Orin NX 16GB has 100 Sparse/50 Dense TOPS. By comparison, both the 8GB and 16GB Xavier NX have a 21 Dense TOPS score and a 10-20W TDP.

The Orin NX offers “almost 5X the performance of the Jetson Xavier NX, or up to 3X performance at the same price,” claims Nvidia. The 3X claim appears to refer to the 8GB model, which costs the same as the 8GB Xavier NX. The 16GB model costs $100 more than the 16GB Xavier NX. The Orin NX is also claimed to offer up to 3x the performance of the AGX Xavier.

Here is an updated comparison of the Nvidia Jetson lineup, ranked in order of ascending processing power:

  Jetson AGX Orin Developer Kit

The Jetson AGX Orin Developer Kit ships with the AGX Orin with 32GB RAM for $1,999. The kit will not be available with the 64GB RAM model or the Orin NX, but the system can emulate these other Orin modules, says Nvidia. Applications include advanced robotics and edge AI applications for manufacturing, logistics, retail, service, agriculture, smart city, healthcare, and life sciences. Jetson AGX Orin Developer Kit (click images to enlarge) The 110 x 110 x 71.65mm dev kit features a PCIe x16 slot that supports PCIe Gen4 x8. There is also a standard RJ45 port that can be used for a fourth 10GbE.

The kit is further equipped with a microSD slot and M.2 M-key with PCIe Gen4 x4 for NVMe storage. An M.2 E-key slot is preloaded with 802.11ac WiFi/BT, which is not available on the AGX Orin module itself. Other features include 16-lane MIPI-CSI-2 and DP 1.4a (+MST).

There are more USB ports than originally revealed. You get 4x USB 3.2 Gen2 ports, 2x of which are Type-C with USB-PD support, as well as 2x USB 3.2 Gen1 and a micro-USB 2.0 port. The kit ships with Nvidia JetPack with Ubuntu Linux.

The Jetson AGX Orin Developer Kit supplies a 40-pin GPIO header (I2C, GPIO, SPI, CAN, I2S, UART, DMIC), as well as headers for audio, JTAG, fan, and “automation.” You also get a DC jack, an RTC with a battery connector, and power, force recovery, and reset buttons. A thermal solution is included.

The Jetson AGX Orin Developer Kit is available from Arrow at $1,999 with a 32GB RAM Jetson AGX Orin. The lead time is 8x weeks. Prices for the AGX Orin and new Orin NX, which will ship in volume in Q4 2022, are listed farther above.

More information may be found in Nvidia’s AGX Orin dev kit announcement and AGX Orin dev kit and Orin NX product page, as well as this Nvidia blog post on the Orin NX. More on other Nvidia GTC announcements may be found in this wrap-up of Nvidia CEO Jensen Huang’s keynote, as well as other links farther above.