Edge AI: platform comparison

Anyone building an industrial product with vision or edge AI runs into an NVIDIA Jetson early. The developer kit is easy to get, CUDA and TensorRT are familiar, and a first model runs quickly. Qualcomm modules, led by the Dragonwing IQ-9075, now match the smaller Jetson modules on the headline number, and often pass them.

The decision has become a volume question. At low volumes the ecosystem outweighs the difference per module. At higher volumes Qualcomm’s extra engineering pays for itself through the bill of materials, and often through power and cooling as well. If the application fits a small industrial NPU, both modules are more than the product needs.

TOPS are not a comparison

A TOPS figure is the vendor’s worked example, not a measurement of your model. NVIDIA quotes Jetson in sparse INT8. Sparse sits at about twice dense, because the chip is scored as if a large share of the weights were zero. Qualcomm leads with the dense number on the IQ-9075: up to 100 dense INT8 TOPS. Read the same way as Jetson, that is about 200 sparse TOPS.

On that scale the IQ-9075 sits above the Orin Nano and the Orin NX, and below the AGX Orin. Jetson Thor belongs in another class. Its headline number is in FP4 TFLOPS, and that figure does not sit next to an INT8 TOPS number.

Module	Headline figure	How that number is counted
Jetson Orin Nano	34–67 TOPS	sparse INT8, Super at the top of the range
Jetson Orin NX	70–157 TOPS	sparse INT8, Super at the top of the range
Qualcomm Dragonwing IQ-9075	100 dense TOPS	INT8, sparse about double
Jetson AGX Orin	200–275 TOPS	sparse INT8
Jetson Thor	2070 FP4 TFLOPS	sparse FP4, a different unit

Compare dense with dense, or sparse with sparse. A higher TOPS number from another vendor is not yet a faster module.

Super mode on an Orin Nano or Orin NX is the same module with a higher power budget. The TOPS number goes up, and the thermal design of the enclosure has to follow.

Whether that number shows up in the product depends on the model, the quantization and the rest of the pipeline. Camera capture, decoding, scaling and drawing the result shape latency at least as much as the accelerator does. How that pipeline fits together is covered in YOLO on an embedded device.

Three families, three software stacks

The module you buy is a SoC plus a software platform. You do not separate the two later.

Jetson comes with CUDA, TensorRT and JetPack. That is a GPU stack with a deep supply of examples, pretrained models and people who have already done it. What you pay for that is a module that draws relatively more power, and an application that over time stays tied to that stack.

Qualcomm comes with the Hexagon NPU and QAIRT, the runtime that gets a model onto that NPU. The IQ-9075 is built as an industrial module: up to 100 dense TOPS, ECC memory, an ambient temperature of -40 to +85 °C, a real-time subsystem alongside, and a stated availability through 2038. Ubuntu and Yocto are both supported. Up to sixteen cameras at once is a different argument from the TOPS number, when the product really has to process that many sensors.

You choose the form factor at the module vendor. Qualcomm itself supplies a soldered BGA module of 65 by 65 mm. Other builders put the same SoC on a format such as SMARC. An Orin NX or Orin Nano sits on a SO-DIMM connector. That difference decides the carrier board, production and how you replace a module in the field.

An i.MX 8M Plus, an i.MX 95 or a TI AM62A is an industrial application processor with an NPU of a few TOPS. The software is Yocto-first. Power stays low, the lifecycle is long, and a quantized model on one camera often fits. Once you need CUDA libraries or tens of TOPS, you are outside this class.

You are choosing a combination of accelerator, runtime, BSP and the people who can maintain that for years.

Low volumes: Jetson, because of the ecosystem

On a first run of tens or a few hundred units, engineering dominates the bill of materials. A difference of a few hundred euro per module disappears next to the hours for the BSP, the camera pipeline, inference and the HMI.

That is where Jetson is the easier path.

• A developer kit is in stock. The application can start while the carrier board is still in design, the same reasoning as with a SOM versus a custom PCB.
• CUDA and TensorRT have examples for nearly every vision model a team wants to try in the first month.
• People with Jetson experience are findable. QAIRT experience much less so.
• Existing CUDA code, a TensorRT engine or a library that only runs properly on NVIDIA does not have to be rewritten.

JetPack starts from Ubuntu. For an industrial product you still want that stack in a reproducible image afterwards, with your own update path. Yocto on Jetson is possible, but it is more work than on a processor whose vendor ships Yocto as the primary BSP. Why that image matters is covered in Yocto vs Ubuntu vs Debian.

Higher volumes: Qualcomm, because of unit price

NVIDIA publishes a guide price from 1,000 units, and that price is already the level a smaller buyer rarely gets under. In autumn 2026 an Orin Nano 4GB sits at 349 dollars, an Orin Nano 8GB at 399 dollars, an Orin NX 16GB at 999 dollars and an AGX Orin 64GB at 2,999 dollars. Those figures change. The pattern stays: the large Jetson modules are expensive, and the discount past 1,000 units is small.

The IQ-9075 has no public price schedule. The module is quoted per project, often with 18 or 36 GB of ECC memory. You set that quote next to the Jetson that can actually carry the application: an Orin NX around 1,000 dollars from 1,000 units, or an AGX Orin at several times that. The difference per unit, times the volume, goes against the extra integration hours. The discount on a Qualcomm module sits in that quote, not on a price list.

A simple worked example

Suppose the Qualcomm module comes out 200 euro cheaper per unit than the Jetson that can do the same compute, and integration costs an extra 400 hours. At 100 euro an hour that is 40,000 euro. The turning point then sits around 200 units. At 2,000 units a year the difference is 400,000 euro a year, before you count cooling.

The real numbers come from the quote and from a measurement of your own model. The principle is the same as the choice between a SOM and your own processor board. Volume changes which cost weighs the most.

At low volumes you pay for engineering and risk. At high volumes you pay for the bill of materials.

Power belongs in the same sum. Inference runs for the whole shift. A module that draws fewer watts for the same frame rate needs a smaller supply, less copper and more often a sealed enclosure without a fan. You do not see that on the TOPS line. You see it in the cost of the device and in reliability on the shop floor.

Two conditions belong with that sum.

The model has to fit the Qualcomm NPU. A network tied to CUDA kernels or a TensorRT plugin does not move because you swap the module. Unit price never catches the extra software then.

And you compare the module that can carry the application, not the most expensive Jetson with the cheapest Qualcomm. An Orin Nano of a few hundred dollars stays cheaper than an IQ-9075 when 34 to 67 sparse TOPS are enough. The price argument starts at the Orin NX and gets sharp at the AGX Orin.

When a small NPU is enough

For many industrial products the application is one camera, a quantized detection model and an HMI next to it. A few TOPS of NPU are enough for that. An i.MX 8M Plus or a comparable processor keeps the power, the bill of materials and the Yocto stack of an embedded Linux product, without CUDA and without QAIRT.

Our X Series is built that way: an i.MX 8M Plus with an NPU on our own carrier, because that class fits an industrial panel PC. Putting a Jetson or a Qualcomm in it would not make the inference better. It would make the module, the power and the software maintenance heavier.

The rest of the field

Rockchip, with the RK3588 as a well-known SoC, delivers a lot of inference for less money than a Jetson. That SoC’s NPU sits around 6 TOPS and has its own runtime, RKNN. BSP quality depends on the module vendor. For a cost-driven product with a shorter life that can be a strong fit. For a device that stays in the field for ten years, that variation in maintenance weighs more.

A separate accelerator, such as a Hailo next to an ordinary industrial SOM, keeps the Linux platform you already have. The AI becomes a second choice, not a reason to move the whole compute platform. That pays off when one model has to run and the rest of the product is a classic embedded Linux application.

Intel with OpenVINO stays relevant when the software already runs on x86 and the enclosure can take that processor’s power. For a compact industrial device without a fan, it is rarely the first choice.

Decision matrix

Factor	Jetson	Qualcomm IQ-9075	Small industrial NPU
Low volumes	Strong fit, ecosystem	Extra engineering	Strong when the application fits
High volumes	Module price stays high	Unit price and power	Often the lowest bill of materials
Software	CUDA, TensorRT, JetPack	QAIRT, Ubuntu and Yocto	Yocto, TFLite or ONNX
Typical application	Heavy models, CUDA libraries, several streams	Many cameras, vision, efficiency	One camera, quantized model, HMI
Industrial	Industrial variants, narrower range	ECC, -40 to +85 °C, long stated availability	Long lifecycle, low power
Dependency	CUDA	Qualcomm runtime	Models stay portable

How we look at it

At Invisto we start from the application and the volume, not from the module with the highest headline number. First we measure whether a quantized model hits the frame rate on a small NPU. When it does, the industrial application processor stays the calmer architecture. When the model or the camera count falls outside that, we compare Jetson and Qualcomm on three things at once: whether the software stack carries the application, whether the module fits the enclosure, and at which volume the unit price earns back the extra engineering.

A prototype may start on the module with the richest ecosystem. The production run does not have to stay there, once the volume and the measurement no longer justify it.