Nvidia’s New Orin Nano Developer Kit: Like a Raspberry Pi for AI
Nvidia’s range of Jetson boards is uncommon raspberry pi Alternative proposal. Nvidia focuses on entry-level AI-based robotics, drones and cameras. Its latest board, the $499 Jetson Orin Nano, packs more processing power while keeping the kit compact.
Jetson Orin Nano improves Nvidia Maxwell GPU’s 128 CUDA cores with 1024 amp based CUDA cores. With additional cores and a new architecture, Orin Nano has up to 80x more AI performance than Jetson Nano. Six Arm A78AE CPU cores provide nearly 7x performance over Jetson Nano. The same AI architecture that powers Jetson AGX Orin modules is used in his Orin Nano, but at a much more affordable price.
Nvidia sent me a sample of the Jetson Orin Nano for review, but due to logistical issues the unit arrived with very little time for a full review. Unfortunately, we were unable to test the main use cases of Orin Nano, inference, and machine learning. Because what I tried with this beta level software didn’t work in my limited time. We plan to publish a full review, including inference benchmarks, in the next few days once we have a new build of Orin Nano’s software.
Please note that the JetPack software provided with the board is a private preview and does not reflect the final software available to consumers. Once the final software release is available, we will provide a full review of Orin Nano, including its powerful AI capabilities.
Jetson Orin Nano Specifications
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 | ||
GPUs | Nvidia Ampere architecture with 1024 Nvidia CUDA cores and | Nvidia Maxwell architecture with 128 Nvidia CUDA cores |
32 tensor cores | ||
memory | 8GB 128-bit LPDDR5 | 4GB 64-bit LPDDR4, 1600MHz 25.6GB/s |
68GB/s | ||
depository | Micro SD | 16GB eMMC 5.1 |
NVMe M.2 via carrier board | Micro SD | |
Power | 7W to 15W (9 to 19V) | 20W (5V max at 4 amps) |
size | 69×45×21mm | 69.6×45×20mm |
Jetson Orin Nano Carrier Board Specifications
Header Cell – Column 0 | Jetson Orin Nano | jetson nano |
---|---|---|
camera | 2x MIPI CSI-2 22 pin 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 | 1x |
USB | 4 x USB 3.2 Gen2 | 4x USB 3.0 |
1 x Type C for debug and device mode | 1 x USB 2.0 Micro-B | |
networking | gigabit ethernet | gigabit ethernet |
RTL8822CE 802.11ac PCIe Wireless Network Adapter | ||
screen | DisplayPort 1.2 | HDMI 2.0 and eDP 1.4 |
GPIOs | 40-pin GPIOs | 40-pin GPIOs |
12 pin button header | ||
4 pin fan header | ||
Power | DC 9-19V barrel jack | DC Barrel Jack 20W (5V Max at 4 Amps) |
size | 100 x 79 x 21 mm (height includes Orin Nano module and cooling solution) | 100 x 80 x 29mm (height includes Jetson Nano module and cooling solution) |
At first glance, the Orin Nano and Jetson Nano look the same. The advantages of the Orin Nano are the fan built into the heatsink and the lack of an HDMI port. The USB-C port replaces the Jetson Nano’s micro USB. The aforementioned fan is whisper-quiet, even when running at full 15W. I ran one of Nvidia’s recommended inference benchmarks, and the fan remained silent unlike the others I tested with the SBC.
reasoning test
Now this section is short and not very sweet. I couldn’t confirm his Nvidia claim that the Orin Nano offers almost 30x the performance of the Jetson Nano (which they hope to improve by 45x).
The root cause of this is the short timescale and private software builds. I wanted to demonstrate the Hello AI World using the Raspberry Pi Camera Module 2, but I ran into a camera issue where the software encoder was not detecting the camera, even though it was listed as compatible. bottom. These issues have been reported to his Nvidia and hopefully future JetPack OS releases will fix these issues.
desktop experience
Running JetPack 5, a custom version of Ubuntu 20.04, 8 GB of LPDDR5 and a 6-core Arm CPU provide plenty of power for typical desktop duties. However, I wouldn’t recommend investing $500 in this board just to use it as a desktop PC.
The first boot was a bit slower than expected, but Nvidia says in their reviewer’s guide that the final production unit won’t have this issue. Another issue we found was that in the preview build he only had 6.3 GB of RAM available. The end user will be able to utilize the full 8 GB with the fix. The Ubuntu experience was pleasant, with minimal desktop customization other than installing tools specific to Orin Nano’s strengths.
Installing Chromium took a little longer than expected. It looks like you installed the browser via Snap, Canonical’s preferred packaging platform. Call us obsolete, but we still love APT a lot.
Once installed, I opened Chromium and went to YouTube to watch some HDR and 4K videos. The first was LeePSPVideo’s HDR video test set to fullscreen and 1440p. Video playback was great, as geek stats showed that a 1440p 30fps video dropped just a few frames.
If I hadn’t used geek-friendly stats, I wouldn’t have noticed at all. The next video, a trip through Costa Rica and its wildlife, played in his 1440p fullscreen, but this 60fps video of him was even worse. About 4% of the frames were dropped throughout the run, mostly at the start of the video. Despite that issue, playback was great.
What the Orin Nano lacks is a dedicated hardware encoder (NVENC). Instead, Nvidia offers a software encoder using his 6-core Arm A78AE CPU. This looks like a downgrade from the Jetson Nano, but perhaps the two extra Arm CPU cores are there to make up for it?
The lack of a hardware encoder also has an impact on how Orin Nano uses the camera. There are two 15-pin CSI connectors on the left side of the carrier board. Compatible with CSI cables made for raspberry pi zeroI tested a simple script to connect a Raspberry Pi Camera Module 2 to CAM0 and record a video. Unfortunately, this was not possible with preview builds of the OS. Even though the Raspberry Pi Camera Module 2’s IMX219 sensor is compatible, I was unable to get an image.
Using GPIOs
Orin Nano’s 40-pin GPIO is on the right side of the carrier board, which is the first problem. Which pin are you connecting to? On the Jetson Nano I printed the board reference as a silkscreen next to the pins.
For the Orin Nano you have to flip the board over and perform a feat of mental dexterity to remember where each pin is. The Python example using BCM mapping in the official tutorial was complicated and I had to decode it further. The Python module is RPi.GPIO, a module familiar to Raspberry Pi fans. Created by Ben Croston, this Python module has powered thousands of Pi projects and quite a few Jetson projects as well. This module has been tweaked to run on Jetson boards, making it more friendly than ever. To avoid pin mapping from the BCM to his BOARD, I chose physical (BOARD) pin mapping despite years of experience teaching Raspberry Pi based content.
It worked and the LED blinked. GPIO pins also provide many of the usual communication protocols. From simple digital IO to UART, SPI, I2C and I2S. Orin Nano’s GPIOs are not the focus of the board, but rather an additional feature for those who want to integrate machine learning with robotics or an array of sensors.
Nvidia’s Jetson Orin Nano developer kit is available now for $499 through authorized resellers.