At every defense tech conference, the same words will be said over and over again: Edge AI, self-driving cars, real-time inference, and combining sensors. The slides look good. The demos are really great. And very little of it shows what really happens when you take that hardware outside and put it somewhere really tough.
We now know how to run AI in a lab with controlled temperatures and humidity. That’s a whole different conversation when it comes to running AI on a platform that is moving over rocks, baking in the desert heat, and drawing power from a source that was never meant to support a supercomputer. And the business world has been avoiding this conversation for a long time.
When it comes to engineering, thermals are the first thing that break. A lot of heat is made by commercial GPUs, and a lot of edge platforms don’t have enough space for the fans, airflow, or heatsinks that are needed to deal with it. A chip slows down when it gets too hot. Inference slows down when it throttles. The mission is harmed when inference takes too long on a real-time defense platform. The effects of that chain of events happen faster than most companies will say in their product literature.
The second wall is power. This kind of platform doesn’t connect to the grid. They can be powered by batteries, the power rails on a vehicle, or generators on board that have strict limits. It’s not up to the designer to choose the size, weight, and power of the SWaP envelope. It’s a rule that can’t be broken, just like the laws of physics. Every extra watt used takes a watt away from the mission’s duration. Someone had to make the case for every extra pound of payload. Most commercial AI hardware fails in this area because it wasn’t made with that math in mind.
Also, the problem with the data pipeline doesn’t get nearly enough attention. A lot of time is spent talking about how accurate models are and how many parameters they have. Not nearly as many people talk about what happens upstream, like the huge amount of EO/IR imagery, LiDAR returns, radar signals, and telemetry that needs to be taken in, synchronized, and processed in milliseconds. Even if a model is perfectly optimized, it won’t help if the pipeline that feeds it is clogged. Because of this, high-bandwidth architectures with dual GPU configurations were created to handle sensor loads that would slow down a normal setup.
Then there’s making it tough. It sounds like a simple task until you understand what it entails. Heat is trapped by protective shielding. Protecting against EMI adds weight. Some parts can’t even be used because of radiation hardening. The mechanical design is changed from the inside out by the needs of the shock and vibration mount. On top of all that, there is security. Edge AI systems are working in more and more dangerous places where threats aren’t just physical. These things—boot chains, firmware, encrypted storage, and zero-trust principles that go all the way down to the hardware—were not just added on later. They are requirements that affect every decision made about the design earlier on.
The defense procurement world as a whole still doesn’t seem to fully understand how much integration risk is behind all of this. Badly matched interfaces and unexpected heat issues during platform qualification can stop a program for months, even if the hardware is solid and the software stacks are well-developed. There is a dollar amount on that kind of schedule slip. Because of this, systems that have already been tested and can work with other systems have quietly become more valuable than the compute numbers on a spec sheet.
It’s interesting that many of these problems keep having small solutions come up. Rugged small-form-factor mission computers, VPX-standard GPU modules, and conduction-cooled boards made for temperatures above 50°C are not the kind of hardware that makes for flashy keynotes. However, it is the equipment that does the work in places where normal AI hardware would stop working. That seems a bit unwilling to change. A big part of the AI industry is trying to get bigger. Someone also quietly solved a harder problem in a two-inch ARM-based board made to last in the desert.
