Edge AI on the Battlefield: Why Pi-Scale Inference Matters for Defence

02 Jun 2026 · 7 min read

A Raspberry Pi 4 Model B. The CM5 with Hailo-8L delivers 13 TOPS at 12 watts, enough for real-time battlefield inference. Photo: Michael Henzler, CC BY-SA 4.0

----|-------------------|-------|---------------| 13|| RPi 5 CPU-only | ~2 FPS | ~8W | 4.0 | 14|| RPi 5 + Hailo-8L | ~30 FPS | ~12W | 0.4 | 15|| Laptop GPU (RTX 4060) | ~800 FPS | ~115W | 0.14 | 16|| Cloud GPU (A100 via SATCOM) | ~5,000 FPS | ~400W + 250ms latency | 0.08 + link dependency | 17| 18|The cloud GPU route delivers peak throughput, until the link drops. And the link will drop. 19| 20|## What Ukraine Proved 21| 22|RUSI's field research drones operating in high-EW zones lose their data link 30 to 60% of the time. Drones that rely on streaming video to a remote analyst become effectively blind during these windows. Those with onboard classification, running on Hailo modules or Nvidia Jetson units, continue tracking and identifying targets even when the operator has no video feed. 23| 24|The lesson has not been lost on the Pentagon. The US Department of Defense's CDAO (Chief Digital and Artificial Intelligence Office) has prioritised on-device AI for GPS-denied and jammed environments through programmes like Project RACE and the Army's tactical cloud initiatives. The direction is clear: push inference to the edge, make it work without continuous connectivity, and accept that the cheapest and most reliable AI infrastructure on a battlefield may not be a server farm but a $70 module on an M.2 hat. 25| 26|## The Hailo Gap 27| 28|Three NPU form factors compete in the sub-15W market, and the gap between them is instructive. 29| 30|| NPU | TOPS (INT8) | Power | Form Factor | Framework Support | 31||-----|---------|-------|-------------|-------------------| 32|| Hailo-8L | 13 | ~2.5W | M.2 Key B+M, HAT | TF Lite, ONNX, PyTorch | 33|| Google Coral TPU | 4 | ~2W | USB, M.2, PCIe | TF Lite | 34|| Intel Movidius Myriad X | ~1 | ~1.5W | USB, M.2 | OpenVINO, TF | 35| 36|Hailo's 13 TOPS is a 3x advantage over Coral and a 13x advantage over the now-discontinued Movidius line. In practical terms, this means the Hailo-equipped Pi runs YOLOv8n at 30 FPS, sufficient for real-time drone video analysis, while Coral struggles at 8 to 10 FPS. On a battlefield where identifying a T-72 in a treeline requires every frame, that gap matters. 37| 38|## Three Deployment Scenarios 39| 40|Disposable sensor swarms. A quadcopter with a Pi + Hailo module running real-time target classification costs under $300 to build. Lose one? Deploy the next. These are not million-dollar Reapers with billion-dollar satellite links. They are consumables. In a war of attrition fought with electronics, the side that fields 10,000 cheap AI sensors against 100 expensive ones wins the information battle. 41| 42|Edge relay for legacy platforms. Older military drones and surveillance sensors were designed before on-device AI was practical. A Pi + NPU module strapped to the data output of an existing sensor feed can add real-time classification without modifying the platform, with no airworthiness certification and no months of integration testing. The Indian Army's existing Heron and Searcher II fleet, for instance, could benefit from this without touching the airframe. 43| 44|Disconnected operations. A soldier in a deep patrol scenario with no connectivity to battalion HQ runs acoustic threat detection, image classification, and SIGINT processing on a device that fits in a cargo pocket. The data is stored locally and uploaded when the patrol returns to comms range. This is not futuristic; it is within the capability of current hardware. 45| 46|## Where This Breaks 47| 48|Edge AI is not a panacea. A Pi-scale device cannot train models; it can only run pre-trained ones. Model updates require physical access or a brief connectivity window. The Hailo-8L's 13 TOPS is insufficient for large transformer models or multi-modal fusion. And running inference on the edge creates a maintenance burden: someone must update the models, manage the storage, and replace failed units. 49| 50|But these are logistics problems, not fundamental barriers. And logistics problems have been solved for every tactical technology before this one, from the radio to the GPS receiver to the night vision monocular. 51| 52|### Beyond Object Detection: What Onboard Inference Actually Enables 53| 54|The argument for edge AI usually stops at "it identifies targets," but the scope of on-device inference goes considerably deeper. The same Pi-class hardware that runs YOLOv8n at 30 FPS can also process acoustic signatures in real time. A three-microphone array connected to a Pi 5 can triangulate gunshot direction of arrival within five degrees of accuracy, classify the weapon type from its acoustic profile, and log the GPS coordinates of the shooter. All of this happens on-device, without any network connection. The soldier gets actionable intelligence in seconds, not in the minutes it would take to relay audio back to a battalion-level fusion centre. 55| 56|Signal classification is another domain where low-power inference punches above its weight. Commercial off-the-shelf SDR receivers paired with lightweight neural networks can identify the modulation type, frequency hopping pattern, and approximate emitter type of an intercepted transmission. The hardware cost is under $200. The processing runs on the Pi. The operator learns whether the signal is a friendly datalink, a Russian R-187 radio, or an ISM-band telemetry channel without ever connecting to a higher echelon's SIGINT pipeline. 57| 58|In practical terms, a single $300 device can serve as an early warning node that simultaneously watches for vehicles, listens for gunfire, and sniffs for electronic emissions. No satellite uplink. No analyst in a bunker 500 km away. The triage happens at the point of collection. 59| 60|### The Model Management Problem 61| 62|The weakest link in edge AI deployments is not the hardware. It is the lifecycle of the neural network itself. A model trained on open-source datasets performs well in controlled environments but degrades sharply when confronted with camouflage patterns it has never seen, weather conditions it was not trained on, or target types that did not exist when the training corpus was frozen. 63| 64|The US Army's experience with Project RACE has made this painfully clear. Models that scored 92% mAP in the lab dropped to 61% when deployed to operational units operating in different terrain and seasonal conditions. The fix is not better hardware; it is a pipeline that allows rapid model retraining and field updates. This is where the Pi-scale approach has a structural advantage over larger platforms. A cloud-connected model update takes weeks of security review, bandwidth scheduling, and interoperability testing. A model update for a tactical edge device can be pushed via a Raspberry Pi in a brief connectivity window over an encrypted USB stick carried by a logistics UAV. 65| 66|The Indian military's procurement culture is not optimised for this cadence. Current DRDO and HAL timelines for software updates on military systems are measured in quarters, not days. Edge AI at scale will require a doctrinal shift in how the services think about software sustainment. The hardware is ready. The organisational process is not. 67| 68|## The Strategic Read 69| 70|For the Indian military, the implication is direct. The MQ-9B SeaGuardian deal, valued at $3.99 billion for 31 aircraft, brings strategic-level ISR. The Archer-NG and Ghatak programmes target stealth combat UCAVs for the 2030s. But between the strategic tier and the stealth future lies a vast operational gap filled by thousands of tactical drones, the kinds that get shot down, crash into trees, and run out of battery mid-mission. 71| 72|These tactical drones need edge AI not as a luxury but as a requirement. The PLA's electronic warfare capabilities along the LAC (direction-finding, signal jamming, GPS spoofing) are well-documented. An Indian tactical drone operating in Eastern Ladakh that needs a satellite link to identify a target will fail when that link is contested. One that processes onboard will not. 73| 74|There is also a cost argument that is rarely made explicit. A single MQ-9B SeaGuardian costs approximately $130 million. A Pi 5 with a Hailo-8L and a good camera costs under $400. The ratio is 325,000 to 1. Strategic assets will always be necessary, but the battlefield is increasingly filled with cheap, attritable platforms that will not get a $130 million data link budget. Edge AI makes them smart anyway. 75| 76|The market is already voting. Hailo's H-8L is shipping at volume, Raspberry Pi CM5 orders are backlogged, and defence contractors are quietly evaluating both. The hardware is cheap, the power budget is manageable, and the operational need is proven. The only question left is whether procurement timelines will match the pace of tactical reality, or whether soldiers will start buying their own. 77| 78|Sources: Raspberry Pi CM5 official, AI Kit Hailo-8L specs, Hailo-8L datasheet, Google Coral products, MLCommons MLPerf Edge, DefenceScoop Pentagon edge AI, RUSI AI in military research, C4ISRNET edge computing 79|