Nvidia’s vary of Jetson boards will not be typical Raspberry Pi options. Nvidia’s focus is on entry-level AI primarily based robotics, drones and cameras. Its newest board, the $499 Jetson Orin Nano ups the processing energy whereas nonetheless conserving the package compact.
The Jetson Orin Nano improves on the Nvidia Maxwell GPU’s 128 CUDA cores with 1024 Ampere primarily based CUDA cores. The additional cores and newer structure implies that the Orin Nano has as much as 80 instances the AI efficiency of the Jetson Nano. The six Arm A78AE CPU cores present virtually seven instances the efficiency of the Jetson Nano. The identical AI structure that powers the Jetson AGX Orin module is used within the Orin Nano, however at a way more inexpensive worth.
Nvidia shipped me a Jetson Orin Nano pattern for evaluation, however resulting from some logistical points, the unit arrived with little time for a full evaluation, so right here I’ll present a short introduction to the board and comment on what I discovered. Sadly, I used to be unable to do any testing of the principle use case for the Orin Nano, inference and machine studying, as a result of what I attempted on this beta-level software program simply didn’t work for me within the restricted time I had it. We’ll publish a full evaluation with inference benchmarks in a number of days, after we’ve hopefully been in a position to get a more recent construct of the Orin Nano’s software program.
Observe that JetPack software program supplied with the board is a non-public preview and doesn’t mirror the ultimate software program that can be accessible to customers. As soon as the ultimate software program launch is made accessible, I’ll present a full evaluation of the Orin Nano, together with its highly effective AI capabilities.
Jetson Orin Nano Specs
| Header Cell – Column 0 | Jetson Orin Nano | Jetson Nano |
|---|---|---|
| CPU | 6-core Arm Cortex-A78AE v8.2 64-bit CPU | Quad-core ARM Cortex-A57 MPCore processor |
| 1.5MB L2 + 4MB L3 | ||
| GPU | Nvidia Ampere structure with 1024 Nvidia CUDA cores and | Nvidia Maxwell structure with 128 Nvidia CUDA cores |
| 32 Tensor cores | ||
| Reminiscence | 8GB 128-bit LPDDR5 | 4 GB 64-bit LPDDR4, 1600MHz 25.6 GB/s |
| 68 GB/s | ||
| Storage | Micro SD | 16 GB eMMC 5.1 |
| NVMe M.2 by way of Provider Board | Micro SD | |
| Energy | 7W to 15W (9-19V) | 20W (Max 5V at 4 Amps) |
| Dimensions | 69 x 45 x 21 mm | 69.6 x 45 x 20 mm |
Jetson Orin Nano Provider Board Specs
| Header Cell – Column 0 | Jetson Orin Nano | Jetson Nano |
|---|---|---|
| Digital camera | 2x MIPI CSI-2 22-pin Digital camera Connectors | 12 lanes (3×4 or 4×2) MIPI CSI-2 D-PHY 1.1 |
| M.2 Key M | x4 PCIe Gen 3 | |
| x2 PCIe Gen3 | ||
| M.2 Key E | PCIe (x1), USB 2.0, UART, I2S, and I2C | 1 x |
| USB | 4 x USB 3.2 Gen2 | 4x USB 3.0 |
| 1 x Kind C for debug and machine mode | 1 x USB 2.0 Micro-B | |
| Networking | Gigabit Ethernet | Gigabit Ethernet |
| RTL8822CE 802.11ac PCIe Wi-fi Community Adapter | ||
| Show | DisplayPort 1.2 | HDMI 2.0 and eDP 1.4 |
| GPIO | 40 Pin GPIO | 40 Pin GPIO |
| 12 Pin Button Header | ||
| 4 Pin Fan Header | ||
| Energy | DC 9-19V Barrel Jack | DC Barrel Jack 20W (Max 5V at 4 Amps) |
| Dimensions | 100 x 79 x 21 mm (Top consists of Orin Nano module and cooling resolution) | 100 x 80 x 29mm (Top consists of Jetson Nano module and cooling resolution) |
At a passing look, the Orin Nano and the Jetson Nano look an identical. What provides the Orin Nano away are a fan constructed right into a heatsink and the dearth of HDMI port. The USB-C port replaces the micro USB of the Jetson Nano. The aforementioned fan is whisper quiet, even after we are working on the full 15W. We ran one in every of Nvidia’s urged inference benchmarks and the fan stayed quiet, not like different followers we’ve got examined on SBCs.
Inference Testing
Proper now this part is brief, and never very candy. Nvidia’s claims that the Orin Nano delivers virtually 30 instances the efficiency of the Jetson Nano (that it hopes to enhance to 45 instances) I used to be unable to confirm.
The foundation causes of this being a brief timescale and the personal software program construct. I needed to display the Hi there AI World instance utilizing a Raspberry Pi Digital camera Module 2, however I bumped into digital camera points which noticed the software program encoder not detecting the digital camera, regardless of it being listed as suitable. These points have been raised to Nvidia, and I hope {that a} future JetPack OS launch will resolve these points.
The Desktop Expertise
Working JetPack 5, a customized model of Ubuntu 20.04, the 8GB of LPDDR5 and six-core Arm CPU present sufficient energy for basic desktop duties. Nonetheless, we wouldn’t advocate investing $500 on this board simply to make use of it as a desktop PC.
First boot was a little bit slower than we hoped, however Nvidia has said within the reviewer’s information that last manufacturing items won’t have this situation. One other situation we noticed was that solely 6.3GB of RAM was accessible within the preview construct. The complete 8GB can be accessible to finish customers by way of a repair. The Ubuntu expertise was nice, with the minimal quantity of customization made to the desktop, in need of putting in instruments particular to the strengths of the Orin Nano.
Set up of Chromium took a little bit longer than we’d have anticipated. It seemingly put in the browser by way of Snap, Canonical’s most well-liked packaging platform. Name us quaint, however we nonetheless have lots of love for APT.
With the set up full, we opened Chromium after which went to YouTube to observe a few HDR and 4K movies. First was LeePSPVideo’s HDR video check, which we set to fullscreen and at 1440p. Video playback was nice, as stats for nerds confirmed a tiny variety of frames dropped for the 1440p 30fps video.
If we hadn’t used stats for nerds, we’d by no means have seen. The subsequent video, a visit round Costa Rica and its wildlife was performed at 1440p fullscreen, however this 60 fps video fared worse. It dropped round 4% of the frames over its complete run, the overwhelming majority being firstly of the video. Regardless of that situation, playback was nice.
Lacking from the Orin Nano is a devoted {hardware} encoder (NVENC). As a substitute, Nvidia provides a software program encoder utilizing the six-core Arm A78AE CPU. This looks like a downgrade from the Jetson Nano, however maybe the 2 further Arm CPU cores are there to make up for it?
The shortage of a {hardware} encoder additionally impacts how we use a digital camera with the Orin Nano. There are two 15 pin CSI connectors on the left aspect of the provider board. These are suitable with CSI cables made for the Raspberry Pi Zero. We related a Raspberry Pi Digital camera Module 2 to CAM0 and examined a fast script to file video. Sadly this wasn’t to be with our preview construct of the OS. Regardless of the IMX219 sensor of the Raspberry Pi Digital camera Module 2 being suitable, we by no means managed to get a picture.
Utilizing the GPIO
Picture 1 of 3
The 40 pin GPIO of the Orin Nano is on the fitting hand aspect of the provider board and right here is our first situation. What pins are we connecting to? On the Jetson Nano we had the board reference printed as a silkscreen subsequent to the pins.
For the Orin Nano, we’ve got to flip over the board and carry out a feat of psychological dexterity to recollect the place every pin is. This was compounded by the Python examples utilizing a Broadcom (BCM) mapping (Raspberry Pi additionally makes use of BCM mappings in all its official tutorials) which required additional decoding. The Python module is RPi.GPIO, a module that Raspberry Pi followers can be intimately conscious of. Created by Ben Croston, this Python module has powered hundreds of Pi initiatives, and fairly a number of Jetson initiatives too. The module has been tweaked to run on the Jetson boards and is as acquainted as ever to those eyes. To get across the BCM to BOARD pin mappings we selected the bodily (BOARD) pin mappings, regardless of our years of expertise instructing Raspberry Pi primarily based content material.
It labored and we had an LED flashing. The GPIO pins additionally present the standard plethora of communication protocols. From easy digital IO to UART, SPI, I2C and I2S. The GPIO of the Orin Nano just isn’t the main focus of the board, extra of an added function for those who want to merge machine studying with robotics or a sequence of sensors.
Nvidia’s Jetson Orin Nano developer package is on the market now for $499 by way of approved distributors.