Just Super! $249 Jetson Orin Nano Super Developer Kit

A Holiday Surprise! NVIDIA Slashes the Jetson Orin Nano Developer Kit Price in Half to $249 – and It’s Faster! Looky here:

Background

NVIDIA launched the Jetson Orin Nano Developer Kit in March of 2023. It is the logical successor to the Jetson Xavier NX, being the little brother to the Jetson Orin AGX. Priced at $499, it served as the testbed for developing not only Jetson Orin Nano but also Jetson Orin NX based systems.

NVIDIA had traditionally sold development kits at a discount to encourage product development around the Jetson SOM. However, after the pandemic, parts became much harder to source and, even for a company as large as NVIDIA, more expensive.

Now, NVIDIA is back at it again! The price of the Orin Nano has been cut in half. There’s an interesting parallel here. In 2014, the 32-bit Jetson TK1 with 192 CUDA Cores was priced at $192. Here we are, 10 years later, and the 64-bit Jetson Orin Nano with 1024 CUDA Cores is $250. Given the inflation over the last four years, this seems remarkable. It feels like the electronics world is returning back to its original crazy self.

Better Performance

The “Super” designation comes from the introduction of a new power mode. The Orin Nano Super has a 25 Watt power mode, which increases performance from between 30 to 70% over the 15 Watt mode. This is accomplished by adjusting three parameters.

• Memory bandwidth: Increased from 64 GB/s to 102 GB/s (Giga Bytes per second)
• Boost CPU frequency to 1.7 GHz
• Boost GPU frequency to 1020 MHz

This is the power draw for the Jetson module itself, the carrier board and attached peripherals have their own power budget. The carrier board can supply up to 45 Watts, providing ample power for the module and peripherals.

Benchmarks

Here’s a side-by-side comparison of the original and new Super versions:

Jetson Orin Nano Super Developer Kit configuration 

Table 1. Jetson Orin Nano Super Developer Kit configuration comparison

Personally, I’m not a big believer in benchmarks. So many applications are compute or memory bound made on assumptions the programmers made building the app. If you are compute bound, it can be much faster. Same with memory speed. But if you’re slow because memory is full, that’s a whole ‘nother kettle of fish. I’ve seen large companies spend huge amounts of time and money trying to get benchmark numbers to be faster. Many times this doesn’t affect real world performance. More importantly, it doesn’t make a better product.

With that said, things like inferencing times can be measured. Faster is better. Here’s a link to the NVIDIA blog post if you really need a benchmark fix.

Jimmy’s Pet Peeves for the Day

When you have a semi-successful video come out and there are several thousand views the first day, you get a lot of comments. I’d guess that 98% of the comments are positive. Positive comments include thoughtful questions, constructive or reasonable criticisms, and expressions of appreciation. I like comments.

That other two percent? Many are just what I’ll call uneducated. There’s always some criticism about pricing and what people believe it ‘should’ be. That’s fair, people are entitled to their opinion. At the same time, it shows a lack of sophistication on how large companies, world wide supply chains, chip manufacturers, board manufacturers, retailers and so on work together.

It also tends to show that they have not worked in hardware or related software companies. There’s almost always a leap to being a consumer victim, which everyone should find unattractive. If a product doesn’t fit your needs, budget, or expectations, it’s best to move on. Here’s an example, but there are many.

One of the comments I read is that NVIDIA was sandbagging with the Orin Nano. The “Super” level of performance was available day one, and it’s just because NVIDIA was greedy that they charged so much money.

As a consumer, that might be a valid belief. A person on the other side of the fence would say it is naive at best, and just stupid at worst.

At its core, NVIDIA is a fabless chip company. Sure, they make a whole bunch of products. But their DNA is chips. Here’s how chips work. Spend more than $10 Billion dollars to develop and bring a GPU type chip to market over the course of two or three years. Design, testing, all the good stuff. How do you design supercomputer level performance? With other supercomputers, of course.

You send your design over to the chip foundry. Eventually you get your test samples back, and they go through a validation process. The first samples are pretty scarce. Now you get them into the hands of the test technicians, application and systems engineers. CUDA just isn’t going to jump on there by itself, you know.

Sometimes there’s a turn or two to fix issues that arise. NVIDIA is better than most at this, and don’t tend to have this happen. Then you’re ready for production. You tell the foundry to turn the spigot on. You don’t tend to get a good yield on the first production runs, it takes a while to ramp up everything and get it going smoothly.

Lower Yields, More Money

When you have lower production yields, the price of the good chips go up. That’s why you see the first chips in use in the upper-end products. NVIDIA is a little unique in that there’s so much in common in their GPU designs across the product line. This allows a better distribution model, data center GPUs, workstation GPUs, gaming cards, down to Jetson.

You’ll note that the actual software is usually still way behind. Typically the chips will have new features, and it takes time to develop software to take advantage of those features. In addition to the libraries we think about, such as CUDA, there’s all the firmware the chips need. That’s one of the reasons that you see a performance increase of over 2X during the first few years that a Jetson is introduced. The new chip features are being integrated, and libraries are being optimized to take advantage of them.

The Jetson chips themselves usually start in the automotive sector, and work their way over to the developer kits over the course of 12-18 months.

Thermals!

If you’re the board designer, and you’re designing a thermal solution, it’s important that an embedded system has plenty of spare thermal capacity. But not too much. This is planned early in the design process, ensuring a theoretical thermal envelope within which the Jetson must operate. So, in some sense, there’s enough “room” to run the faster clock and bus speeds on day one.

That’s a typical partial truth outside people use. The fact is, embedded systems have to work as delivered on the first day. That’s a daunting task in and of itself. The consumer doesn’t take into account how much simulation and testing it takes to actually find the correct settings for optimal performance in a given power budget. If you’re trying to shoe horn that time in with just getting the chip up and running reliably, you are getting on board the pain train my friend.

Also, it’s a system. There’s balances and tradeoffs between how much of the power budget the CPU, GPU and memory should get. Some of the chips main features may not even have been tapped yet, which can affect those calculations greatly. An incorrect power model can create all sorts of issues, I can’t think of many professional users would find the risk acceptable.

Pricing

As for the pricing, there are many people who complain regardless of the price. Putting that aside, the bean counters can usually get really close on the BOM price. NVIDIA has been doing this a good amount of time, so the slope of the production yield to price is relatively easy to guess. However, what’s not easy to guess is what happens when all the suppliers shutdown and parts are in short supply.

This became especially apparent during the pandemic, though such disruptions are not uncommon. There are many times when there’s an issue where you cannot get a vital part. This can be due to a manufacturer having production issues, supplier company troubles, transportation, and so on. This is a global operation after all.

You have another group of issues when you get it out to retailers hands. Retailers aren’t part of the manufacturer, and it’s through agreements that they price products. There’s most likely minimum stock buys, so the retailer usually is carrying inventory. When the manufacturer announces a 50% price cut, significant coordination is required to reconcile retailer inventory and agreements. And the retail side tends to be a whole bunch of companies, spread around the world.

Conclusion

The Jetson Orin Nano Super Developer Kit is cheaper and faster! What’s not to like? At $249, it’s getting back to a much better value proposition. Check it out!