NVIDIA this week gave its Jetson lineup of embedded computing boards an unexpected expansion this week with the announcement of the Jetson Orin Nano 2. The forthcoming entry-level edge AI board is intended to further flesh out NVIDIA’s existing Jetson Orin board, giving customers a more powerful option than the original Nano, but slotting below the all-around more powerful and more expensive Jetson Orin NX.
Altogether, NVIDIA is touting a 2x increase in AI performance for the updated Nano 2, or alternatively the ability to hit a similar performance level at 40% less power. All of which while running the existing Jetson Orin software stack that NVIDIA has built up for the past few years.
Under normal circumstances, the Jetson Orin Nano 2 would be an unremarkable product, as NVIDIA already has several different Jetson Orin SKUs. But digging a bit deeper, there is a lot more going on for this board than the official specifications or even NVIDIA’s press release let on, as NVIDIA has designed a new chip to power this SKU. As a result, there is quite a bit going on under the hood that makes the Jetson Orin Nano 2 a notable product release from the company.
NVIDIA Jetson Orin Nano 2 #
Diving right in to the specs, here is a high-level look at the specifications for the new Orin Nano 2, and the other Orin products.
NVIDIA Jetson Orin Lineup | |||| Orin Nano (Super) | Orin Nano 2 | Orin NX | AGX Orin | | CPU Cores (Cortex-A78) | 6 | 8 | 8 | 12 | CUDA Cores | 1024 | 1024? | 1024 | 2048 | Tensor Perf(INT8 Sparse) | 67 TOPS | 78 TOPS | 77 TOPS | 156 TOPS | Memory | 8GB | 8GB | 16GB | 64GB | Memory Bandwidth | 102GB/sec LPDDR5-6400 | 120GB/sec LPDDR5X-7500 | 102GB/sec LPDDR5-6400 | 205GB/sec LPDDR5-6400 | TDP | 7W – 25W | 15W – 40W | 10W – 40W | 15W – 60W | Release Date | Q2’2023 | H1’2027 | Q1’2023 | Q1’2023 |
Compared to its original predecessor, the Nano 2 offers a slight hardware increase across the board. Another two Arm Cortex-A78 CPU cores have been added, bringing it to 8 CPU cores, and the overall (theoretical) tensor throughput of the GPU has been boosted to 78 TOPS of spare INT8 math. Meanwhile the 8GB of LPDDR5-6400 memory has been swapped out for faster LPDDR5X-7500.
Based on these specifications alone, it would be reasonable to assume that NVIDIA has just developed another SKU based on their existing Orin silicon, creating a SKU that sits between the Nano and NX. However, looking deeper at the specifications and NVIDIA’s performance claims, things start to break down. How is the 78 TOPS Nano 2 twice the performance of the 67 TOPS first-gen Nano? And how has NVIDIA been able to add support for LPDDR5X memory, which is an adjacent but distinct superset of LPDDR5, to their Orin silicon almost half a decade after it was developed?
The disparity arises because the silicon behind the Orin Nano 2 is in fact new silicon altogether from NVIDIA. As outlined by NVIDIA in a separate press briefing for the product launch, the hereto-unnamed chip is still an Orin architecture design combining Arm Cortex-A78 CPU cores with an Ampere architecture GPU, but with new performance enhancements not found in the original Orin silicon.
Specifically, NVIDIA has done a couple of things to augment the new Orin (at least, what they are disclosing at this time). The first, of course, is adding LPDDR5X support, making this the first Orin product to get support for the faster memory speed grades. The second aspect, meanwhile, is more interesting and nebulous: NVIDIA has made “architecture and microarchitecture” improvements to the Ampere GPU in the new Orin to improve the real-world achievable performance of its 32 tensor cores in AI workloads. The peak theoretical performance may have only improved by 16%, but NVIDIA says that the amount of real-world work that can be extracted from the hardware has doubled – essentially capturing a hefty amount of performance from the silicon that the original Orin could not.
Whatever the case, these changes go beyond just adding more (or wider) functional units, as these would show up in the top line TOPS figures. So it will be interesting to see just what NVIDIA has been up to once the hardware ships and the public documentation is updated.
These performance improvements have also pushed the power/performance curve out for the Jetson Orin Nano 2 versus its predecessor. In 15 Watt mode, NVIDIA says it can deliver the same amount of performance as the first-generation Nano in 25 Watt mode – which is where NVIDIA’s 40% power reduction claim comes from. Alternatively, when running in the new 40 Watt mode, the Nano 2 can deliver twice the performance of the Nano 1 at its peak 25 Watt mode. The higher power mode means that in practice, NVIDIA is fueling some of their performance gains through higher TDPs (Orin NX also goes to 40W), but the overall performance improvement of the Nano 2 (and specifically its ability to hit similar performance at meaningfully lower TDPs) underscores the fact that some significant architectural efficiency improvements had to occur to hit these performance and power numbers.
The new Jetson Orin Nano 2, in turn, will be used to broaden the performance range for NVIDIA’s entry-level AI modules. The Nano 2 will be joining NVIDIA’s original Nano as their entry-level products, each with 8GB of memory. NVIDIA pitches this range at a pretty wide variety of products, covering everything from classic robotics to drones and computer vision systems.
Meanwhile the Orin NX and Orin AGX products will remain as NVIDIA’s mainstream Jetson products, with the latest Thor-based Jetson T-series sitting further above that. Not counting the different memory capacities, this means that the contemporary Jetson hardware stack is now comprised of 9 different products, a deeper stack than ever before.
Filing the Gap Left by Atlan? #
The new Jetson boards aside, the news that NVIDIA has developed a new Orin SoC opens up a wide range of interesting possibilities. And it comes with some significant repercussions as well.
First and foremost, NVIDIA has never refreshed a Jetson Nano product in this fashion before. There have been other generations with multiple Nano SKUs, but they were always based on the existing SoCs initially created for that generation of hardware.
For the Orin generation in particular, this has meant that NVIDIA has stretched the 12 CPU core + 2048 CUDA core chip down to a Nano board with half as many cores (and half as wide of a memory bus) in use, and a whole bunch of other hardware such as the dedicated Deep Learning Accelerator (DLA) hardware going unused. This makes the original Orin a relatively expensive chip to produce in as much as NVIDIA is paying for a lot of functionally dark silicon on Samsung’s 8nm process node. One of the biggest outstanding questions about the new Orin chip, then, is whether this has been designed as a cost-reduced version specifically for the Nano market. Did NVIDIA tape out an Orin-architecture chip with just 8 CPU cores and 1024 CUDA cores to start with, reducing the chip to just what the Nano market needs – essentially crafting an Orin Lite? Looking in from the outside, this has no obvious answer; but looking at the new Jetson hardware stack, there is not much room to squeeze in something after the Orin NX, a part that NVIDIA is already committed to producing through 2032.
What is interesting about this development though is that it comes about 4 years after NVIDIA canceled Atlan, the Arm SoC that was meant to succeed Orin in the 2024 timeframe. Since then, NVIDIA has covered part of the gap with the newer Thor SoC, but the cost and size of Thor has thus far made it a poor choice to use in a new Nano product. Absent Atlan, that leaves Orin to pick up the slack. And 4 years is more than enough time time to develop another Orin chip in reaction to Atlan’s cancelation.
It is curious, however, that NVIDIA would mint a new AI chip based on the Ampere architecture. While it was fine for its time, Ampere lacks support for FP8 and FP4 precisions; INT8 is as low as it gets. FP4 support alone brings the potential for twice the tensor throughput as FP8/INT8, so that would be very desirable to have in a new budget edge chip. That said, Orin IP is readily available for Samsung’s 8nm node, while Thor’s IP is all centered around TSMC 4nm/3nm, node families that are already in high demand for other AI products.
Final Words #
Between the expansion of the Jetson Orin lineup with a newer and significantly faster Nano embedded computing board, and the fact that NVIDIA has designed a new Orin chip for the edge AI module, this is one of the clearest signs yet that that NVIDIA’s edge/robotics efforts have finally grown to the point where it has become a large and self-supporting market segment for the company. Producing a new Orin chip means that NVIDIA expects to be able to recover the costs of its development (and then some) over the lifetime of the chip. This is something that past generations have accomplished by using NVIDIA’s mobile silicon in a broad spectrum of products up to and including their DRIVE automotive systems; a mold the newer Orin chip breaks.
Overall, the entire industry has been pushing the edge/robotics/physical AI angle hard this year. A new Jetson Orin Nano is one more piece of evidence about NVIIDA’s ambitions for the entry-level edge AI market, as well as a sign of how big of a market they expect it to become.
Wrapping things up, as is typical for NVIDIA Jetson announcements, the hardware is being announced several months ahead of availability. The Jetson Orin Nano 2 will be released in the first half of 2027. Pricing has not been disclosed.