The NVIDIA Jetson AGX Orin Developer Kit is now available for purchase!
Big Picture
▪ Up to 275 INT8 TOPS powered by Ampere Tensor Cores GPU + DLA ▪ 12x A78 ARM CPU
▪ 32 GB memory, 204 GB/s
▪ $1999
▪ Available Now
Who is it for?
As you can see from the price, this Jetson is targeted towards the professional development crowd. The dev kit can be set up into different configurations to emulate the different Jetson Orin modules that will be available in Q4. Each of these modules have different configurations with regards to the amount of power they consume, the number of CPU cores, and the number of GPU cores. Good stuff!
Check out the official NVIDIA Jetson Orin page for ordering info.
There are now over 1 Million Jetson developers!!!!
Excuses, Excuses …
Normally on a big release like this, NVIDIA provides JetsonHacks with an early production version before the release. However this year there was a little issue. China initiated a lockdown right when the Orin was on its way to JetsonHacks world headquarters. The AGX Orin is now sitting at an airport in Hong Kong waiting. My guess is that if they can convince it to leave the airport bar, it will get on the next plane to its new home. When we get the unit, we’ll go deep dive on it!
We’ll just drop a little info here, and wait for when we get our greedy little mitts on it to get the real info which we will share here.
The Software
The Jetson AGX Orin will run JetPack 5.0, which brings Ubuntu 20.04 with LTS Kernel 5.10. Some of the other highlights include CUDA 11 and new version of cuDNN and TensorRT. This is currently scheduled to be released March 30, 2022. It should run on both Orin and Xavier architectures.
One of the big announcements is that the NSight Tools Suite can now run natively on the Jetson Orin! That means the large number of professional level developer tools, such as the CUDA profiler, can now run natively, no host machine needed!
Jetson Orin Modules
The configuration and pricing for the Jetson Orin Modules has been announced. The modules will begin shipping in Q4, 2022. There are two configurations of both the AGX formats and NX formats.

Jetson AGX Orin 64GB
275 Sparse|138 Dense INT8 TOPS
15W to 60W
$1599 (1KU+)
Jetson AGX Orin 32GB
200 Sparse|100 Dense INT8 TOPS
15W to 40W
$899 (1KU+)

Jetson Orin NX 16GB
100 Sparse|50 Dense, INT8 TOPS
10W to 25W
$599 (1KU+)
Jetson Orin NX 8GB
70 Sparse|35 Dense, INT8 TOPS
10W to 20W
$399 (1KU+)
Jetson AGX Orin Developer Kit Specifications
JETSON AGX ORIN MODULE | |
AI Performance | 275 TOPs |
GPU | NVIDIA Ampere architecture with 2048 NVIDIA® CUDA® cores and 64 tensor cores |
CPU | 12-core Arm Cortex-A78AE v8.2 64-bit CPU 3MB L2 + 6MB L3 |
DL Accelerator | 2x NVDLA v2.0 |
Vision Accelerator | PVA v2.0 |
Memory | 32GB 256-bit LPDDR5 204.8 GB/s |
Storage | 64GB eMMC 5.1 |
Video Encode | 2x 4K60 | 4x 4K30 | 8x 1080p60 | 16x 1080p30 (H.265) |
Video Decode | 1x 8K30 | 3x 4K60 | 7x 4K30 | 11x 1080p60 | 22x 1080p30 (H.265) |
Refer to the Software Features section of the latest NVIDIA Jetson Linux Developer Guide for a list of supported features
REFERENCE CARRIER BOARD | |
Camera | 16 lane MIPI CSI-2 connector |
PCIe | x16 PCIe slot: Lower x8 PCIe Gen4 |
RJ45 | Up to 10 GbE |
M.2 Key M | x4 PCIe Gen 4 |
M.2 Key E | x1 PCIe Gen 4, USB 2.0, UART, I2S |
USB Type-C | 2x USB 3.2 Gen2 with USB-PD support |
USB Type-A | 2x USB 3.2 Gen2 2x USB 3.2 Gen1 |
USB Micro-B | USB 2.0 |
DisplayPort | DisplayPort 1.4a (+MST) |
microSD slot | UHS-1 cards up to SDR104 mode |
Other | 40-pin header (I2C, GPIO, SPI, CAN, I2S, UART, DMIC) 12-pin automation header 10-pin audio panel header 10-pin JTAG header 4-pin fan header 2-pin RTC battery backup connector DC power jack Power, Force Recovery, and Reset buttons |
Dimensions | 110mm x 110mm x 71.65mm (Height includes feet, carrier board, module, and thermal solution) |
Conclusion
These puppies will be in high demand. Best be getting one while the getting is good.
26 Responses
Not available in the United States
Phil has been working through this issue. The AGX Orins Dev Kits are available to order currently through Arrow: https://forums.developer.nvidia.com/t/orin-developer-kit-availability-in-usa/209126/8
In terms of sheer horsepower, both CPU and GPU, how will the $400 Orin development system compare to the $500 nvidia 2070 Super video card, with it’s 2,000+ Cuda cores? for that matter, how will the $2000 Orin compare, since it only has 2,000 cuda cores???
And if we’re lucky, the 4070 (Super?) video card due out soon, compare? Yes, I know they are made for different markets but I need to know about the sheer horsepower question.
I also heard a rumor that the Jetson Nano’s are no longer produced! Tell me it ain’t so, please!
$1999US? I saw $1499US. But even at $1499US it’s not price-performance competitive with an RTX 3050 laptop. You would only buy one of these if you had a business case for high-capacity real-time video analytics that required the special-purpose video processors on the Orin.
I haven’t heard anything about the Jetsons not being produced, but it does appear they are frozen at the current Jetpack software level, while the Xaviers will support the new Jetpack. And I don’t know of any legitimate non-scalper sources of AGX Xavier or Xavier NX developer kits.
The price for the Dev Kit is $1999 USD. The modules that are being released (AGX in October, NX later in Q4) range in price from $399 to $1599. To be clear, the Dev Kits are not laptop competitors. For people who can use laptops for their application solutions, that’s a great way to go. Especially for something that is low volume.
Jetsons are developer kits meant for people who are developing applications in both hardware and software who need embedded hardware in their designs. For example, NVIDIA has a division that works on the Drive product, which provides infotainment and autonomy elements to the automotive industry. The current Drive product uses two Orin processors. There are over 130 different companies that integrate the Jetsons in their products, and over 6,000 companies that are using Jetsons with their products. Robotics and Healthcare carve out a good slice of the Jetson market, along with video analytics on the edge.
The quote “market” for the AGX Orin dev kit is for developers who need to build new hardware devices or develop software for the new devices. You can configure the dev kit to mimic any of the module configurations. The professional Dev Kits are not designed to be integrated into a product, but rather provide a reference design for people that are designing products. Note that the $1999 price is less than the original AGX Xavier $2499 at its introduction 4 years ago. The AGX Xavier is now around $699 if memory serves.
The entry level Jetson Nano is a different case, which is for makers and people who want to get their feet wet in the embedded space. The pricing of the product line reflects that.
I believe that the strategy that NVIDIA has taken for the short term is to provide available module production to their Jetson Hardware Partners There are many configurations of carrier boards and Jetson modules that you can buy. The pricing is more expensive, but at least it’s available from approved members of the ecosystem. In part this is because NVIDIA has historically subsidized the development kits, making the cost less than the price of the Jetson module to the developer. While it’s a good deal, I think there’s some sticker shock when you go to build your Jetson based product and realize how much more it ends up costing.
You are correct about JetPack. Jetsons earlier than the Xavier will not be able to use JetPack 5. JetPack 4 is now feature complete, though there will be updates from time to time for security and bug fixes. JetPack 5 is a big deal, new boot loader, Ubuntu 20.04, new kernel and so on. The Nsight tools will run natively now on the Xaviers and Orins, which makes sense because the newer Jetsons have desktop performance. This is a welcome addition for people like me who despise the cross development process.
That’s an apple and oranges comparison. The 2070 have a Turing architecture, the Orin is Ampere. There are many ways to guesstimate what the performance will be. Turing is 12nm, where the Orin is a 8nm process. Which means Orin has about 2x the number of transistors. But Orin also has a CPU complex (between 6 and 12 cores), which the discreet GPUs don’t have. The Orins also have microcontroller cores for controlling GPIO and audio. I don’t know specifically, but I would guess the discreet GPUs only use that feature for audio processing. The Orin has a power budget that must work in < 60W; a discreet GPU is hundreds of watts. Plus you have the myriad of side car processors, such as the DLA units and Tensor cores.Also, if you compare an Orin with it's contemporary Ampere architecture RTX 3070, the 3070 has a bigger chip die with 17.4B transistors versus the Orins 13B or so. The 3070 has twice the memory bandwidth. For many applications, memory bandwidth is a performance bottleneck. The discreet GPUs have specialized memory, where as the Orins have a unified memory architecture which shares the GPU memory with the CPU. As a side note, the new Macintoshes also use a unified memory architecture now.I've talked to the people at NVIDIA who matter, and they have said that the Nanos are still in production. But like everyone else, they are experiencing component shortages. The industry analysts say that this probably won't clear up anytime soon. Everyone is scrambling for alternatives. But these are physical products and take time to engineer and bring into production. You will guess that the China Covid lockdowns that are happening right now, and you are right. It's relatively easy to stop everything, getting it started back up is a whole different beast.So what does all that mean? Nothing in particular. It all depends on the applications that you are running as to what type of performance you can expect. One of the main ideas behind Orin is that you can run multiple machine learning models concurrently. On a robot for example you can run vision recognition, obstacle detection, and path planning models concurrently. All this without having to drag a Tesla sized battery along with you.
$2000 was list and the $1500 was for qty 1000, AIUI.
I’m asking because I want so badly to experiment with several Nano’s/Orin’s ($400 qty 1000?) for playing with a cluster of these guys and (probably) need the 2070 Super or 4070 Super as controller for the cluster. I think that will be tons-o-fun. We gotta find our fun somewhere. 🙂 Can’t dance and it’s too wet to plow…
I’m hoping to make a (small) cluster using Julia language with it’s in-built messaging capabilities. Julia also makes it a TON easier program the cuda cores. Assuming I can ever get it to run on the Jetson family. If not, I can’t afford 5 or 6 desktop video cards…
Seeed used to have 4-board Xavier NX clusters for sale but the Xavier NX is unobtainium. If I had the budget for that many 32 bit FLOPS I’d simply buy A100s. 🙂
Incidentally, my application is audio processing – algorithmic composition and digital sound synthesis. I have AGX Xavier, Xavier NX and 4 GB Nano developer kits already plus a 1650 Ti laptop and an RTX 3090 desktop. I’m pretty confident I can develop software on my existing Jetsons that will run on any Orin module and on any WSL 2 / NVIDIA system. That’s where I find my fun, anyhow.
Of course. I’d still like to play with it, though.
Ed, have you looked at the Teensy and it’s audio board? Cheap as dirt but supposedly the specs are audiophile quality. I have both but not tested them that way, yet.
I have a Dirtywave M8 tracker, which uses the Teensy, and four Teensys. I also have an Electro-Smith Daisy Pod and Field and two Bela.io systems. None of these are capable of running FFTs at even a Jetson Nano speed, though. 🙂
1999$ price is insane.
Insane good, or insane bad? The AGX Xavier Developer Kit was $2499 on introduction in 2018. The target market for the AGX Orin is for the early adopters who will be building products and the ecosystem for the Orin module family available later this year. The modules will start at $399. The Orin delivers a large performance boost (~ 1.7x CPU, 5-6x GPU) over the equivalent Xavier in AGX or NX form factor. If you are comparing this to a Jetson Nano, then yes that’s a shocking price. If you are using it for product development, that’s certainly a reasonable price.
Insane-ly bad. Thanks for pointing out that old AGX Xavier was 2500$, I wanted upgrade over Xavier NX kit which costed same as module, therefore I was quite shocked to see the price. I get that price reflects engineering efforts etc, but for that price I’ll upgrade my macbook first 🙂 All that talk from Nvidia supporting many developers, I guess Amazon can buy millions of them for all the interns and not care. Also Turtlebot 4 is coming out, so maybe invest there. It would be nice to have some competition to Jetsons, lot of people are using RPi 4 for basic ROS stuff, but it can’t run NN for vision etc. Hopefully Nvidia won’t blame Covid and war in Ukraine for high prices in 2030s.
Not quite sure what you are saying here. Are you against the introduction of new product lines?
I suggest that a MacBook and a Jetson Orin Dev Kit aren’t quite the same thing. I would think upgrading the MacBook would be a good bang for a discretionary spending buck. iRobot has a new Turtlebot base that you can put a RPi, a Jetson Nano or Jetson NX on. The Orin Dev Kit ain’t about any of that.
I do not know what you mean by the term blame in this context. The effects of Covid on the worlds production and supply chain are well known, and certainly effects companies other than NVIDIA. I’m also not sure what you mean about the war in Ukraine. As I’m sure you know, Ukraine produces ~70% of Neon in the world. Neon is currently essential in chip manufacturing. On the surface, I’m wondering how that could NOT effect prices. Does that carry in to the 2030s? Dunno, maybe everything calms down, or people come up with a work around so they don’t need Neon anymore.
People anthropomorphize companies. That is, they assign human traits to them. While most companies are comprised of people, they are not people. Companies are legal entities. They can’t “care” or “blame”. For example, China does a COVID lockdown, manufacturing and shipping stops in the lockdown area. Companies have their products being manufactured there. A company is obligated to state that they are experiencing an issue. Typically they attribute the issue to a cause. Now you can call that “blame”, but it should be a statement of fact if the company is acting in good faith.
If the company is not acting in good faith, then more than likely the shareholders get really upset as well as company regulators and typically the government. Not to say that all companies are good, or act properly all the time.
I’m also don’t know what you mean about Amazon buying “millions of them for all the interns”. Do you believe that NVIDIA is actually producing Jetson Dev Kits and making them unavailable for people to buy? That there is some secret railroad that sends them to other nefarious groups like Amazon? It’s a great story, I would love to hear it!
For almost any new chip design, the yield is relatively small. Typically there’s a pretty low limit as to how many you can produce until you get yields up, and fix the inevitable errata. Compare the Tegra Orin to the Tegra X1. It took a few years before the Tegra X1 ramped up production. The NVIDIA Shield set top box and Nintendo Switch are just some of the products built around that chip. In part, that’s why the Jetson Nano exists at it’s current price point.
Once the yields start coming up and you build out the rest of the production lines to build something that works, then the costs come down. The Orin is at the very start of this process. They aren’t making millions of Orins, but they are making enough to get to the early adopters on developer kits which will in turn allow people to build the ecosystem around the chip. And they are expensive because they don’t make very many of them in chip making terms.
Wow
“NVIDIA Ampere architecture with 2048 NVIDIA® CUDA® cores and 64 tensor cores”
That’s pretty close to an RTX3050.
If seems that it is quite capable of running the TAO toolkit and Isaac Sim replicator.
Now I know that is not Nvidia’s primary goal with the AGX Orin
But if it can run those 2 pieces of software in addition to all the Horsepower it’s capable of putting out, $1999.00 is a steal for that kind of capable hardware.
Cant wait to get my hands on it and take it around the block and see what it can do.
A couple of things to remember. The RTX3050 has more memory bandwidth and a bigger chip die. The Orin has a 12 core CPU complex (3 blocks of 4 CPUs). Also the Orin limits to 60W, the RTX3050 uses hundreds. I would not expect equivalent performance, but certainly I would expect a desktop like experience. They have ported over the Nsight tools, so it should be much more comfortable to develop on. Thanks for reading!
Agreed
I just got the pre-order availability email from Seeed – they have it at $2250US. This is not a hobbyist unit by any stretch of the imagination. If you don’t have a legitimate expectation of a revenue stream from working with it, you should find another way to build your project.
I totally understand the global economic and marketing factors behind NVIDIA’s pricing decision on the AGX Orin developer kit. But given the lack of existing stock or a commitment to restock the AGX Xavier and Xavier NX developer kits, I think NVIDIA have said they’re no longer interested in the hobbyist, and that bothers me a lot.
You guys crack me up. This roadmap with the Orin has been publicly available for several years. There is a range of Jetson products, from Nano to now Orin. While you are free to build your outrage over non-existent slights, your conclusions don’t make sense. Next year, we’ll see the next Nano. Does that mean that then NVIDIA will no longer interested in the professional market, and is exclusively focused on the hobbyist and education markets instead?
If you have looked at the spot market for components recently, you will find that LPDDR4 memory is very tough to even find, let alone get for a decent price. Along with various other glue chips, like Ethernet controllers. Even commodity STM32 microcontrollers are running $160 per, for a part that normally costs ~$2.
If you follow along in the Raspberry Pi community, you see that they are experiencing the same issues. Are they no longer interested in the hobbyist and educational markets too? In comparison to the Jetson Nano, RPi runs huge numbers. And the bigger contracts usually get serviced first, so depending on the supply lines it will take time to get everything straightened out again. It was almost unimaginable in mid ’20 when people were saying that it wouldn’t be straightened out until mid ’22, early ’23. That’s why everyone is scrambling to change their designs to work around what is sure to be lingering chip shortages. But that takes time.
One of the ways that Orin works around some of these issues is by using newly introduced parts, such as LPDDR5 memory. Older designs cannot accommodate these chips. But you have to remember that NVIDIA has been working on Orin for 3+ years. It’s not just something that you can change overnight.
To be clear, I’m not a NVIDIA fan boy. But I have worked in the computer industry for a very long time. Anyone who have been around for more than a decade remembers that there are chip shortages from time to time. Those shortages are generally relatively short, and you could generally buy yourself out of them. However, this one is much more difficult because of its randomness and severity.
For the Orin in particular, I can tell you kind of what happened. A week ago, shipments were supposed to start. On the Friday evening before, China put a Covid lockdown on Shenzhen and Beijing. China locks down HARD! So guess what, the shipments didn’t happen and the factories closed. In the US, Arrow is the first distributor scheduled to get product. It appears as Arrow’s policy is not to sell products that are not in stock, and not allow back ordering (at least for this particular product).
By late last week, China opened up some of their economic zones so that they could build and ship products. I think you can imagine an optimist thinking that they would get products. Marketing people all have that gift. But you can also imagine that others might make a guess and say that give it a month or two. A better answer is that people don’t know.
NVIDIA isn’t the only one affected of course. Apple had a roll out for several of their new products a couple of weeks ago that ran into the same problem. Of course Apple is in a whole other league with regards to supply chain and logistics. They have an entire city that makes 500K IPhones a DAY! Everything just stopped.
Could NVIDIA communicate all of this better? Sure. But that’s a lot easier to say than do. If you come out every week and say “I don’t know”, is that helpful? They appear to have taken two tacts. First, supply available modules to their partner network. Many of their partners sell carrier boards, some of which are the same as the dev kits. That helps with supply, and helps keeps the partners a little happier. They’re suffering through this too, remember. The second is that NVIDIA says “We’re working on it”. My guess is that’s way truer than it sounds, and that the supply folks are not having a good time.
“…they’re no longer interested in the hobbyist…” <—- That is INFURIATING!
And also not true. But feel free to work yourself up about speculation.
And to think, I had only recently begun forgiving them for the TX2!!!
In my humble opinion I think that the
Jetson Orin NX 8GB
is going to be the sweet spot for hobbyist and production developer’s.
$399 is within reach of hobbyist who have out grown the Jetson Nano an Xavier NX.
Especially if you utilize deepstream.
It should have no problem running YOLO 4 at well above 40FPS.
That’s a good observation. The NX versions have 1024 CUDA cores with 6 or 8 CPU cores. I think the intent is to give people the capability to run multiple DL models at once. The smart people will be leveraging the NVIDIA models that they’ve been building on over the last few years. So you would expect something like RIVA running simultaneously with a domain specific model. It’s always comforting to know that you can go bigger/faster if necessary.