The convergence of highly capable edge AI models and advanced commercial-off-the-shelf (COTS) edge AI accelerators is reshaping how computation is deployed across defense, aerospace, and commercial platforms. Mission-critical decisions increasingly must be made at the edge, onboard vehicles, satellites, and infrastructure nodes, where latency, connectivity, and power availability are constrained.
Edge AI is also beginning to incorporate compact generative approaches to support greater autonomy. This shift is accelerating the adoption of COTS edge AI accelerators as an alternative to centralized processing and legacy radiation-hardened compute architectures.
From Centralized Processing to Edge Intelligence
Traditional space and defense systems relied on centralized onboard processors or ground-based compute for data-intensive workloads. That model is increasingly misaligned with operational realities. High-resolution sensors generate more data than can be economically downlinked, contested or intermittent communications limit reliance on reach-back processing, and general-purpose GPUs often exceed acceptable power envelopes for embedded platforms.
Purpose-built edge AI accelerators, typically ASIC-based and optimized for inference, offer a more effective approach. These accelerators deliver high performance per watt in compact form factors which are well suitable for unmanned systems, tactical vehicles, and small satellites.
Emerging Role of Generative AI at the Edge
Alongside discriminative inference, compact generative AI models are beginning to gain traction in edge-constrained defense, space, and infrastructure systems. Rather than large conversational models (LLMs), these deployments focus on smaller generative architectures optimized for synthesis, reconstruction, and prediction. Typical requirements include batch-size-one execution, mixed-precision support, high memory bandwidth, and deterministic latency under tight power envelopes.
Representative use cases include image enhancement for degraded sensing, predictive “what-if” analysis for system operations and maintenance, and synthetic data generation for cyber and RF anomaly detection. Example edge-relevant generative models include SPAN and XLSR for super-resolution and denoising, and compact language models such as Llama3.2 1B/3B and Gemma3-1B for structured sequence generation and on-device summarization.
Radiation Testing and Implications for COTS Adoption
Space is a frequently cited domain for advanced on-platform autonomy, yet traditional reliance on specialized radiation-hardened devices has limited capability and driven high costs. This has constrained adoption, particularly where commercial economics are a factor.
Recently, NASA conducted a proton and heavy-ion testing of modern edge AI accelerators, the EdgeCortix SAKURA-II, evaluating single-event effects using representative computer vision and classification workloads. The EdgeCortix results are impressive. There were no destructive latch-up events observed through high linear energy transfer (LET) levels and most effects manifested as recoverable single-event functional interrupts or transient inference deviations, many of which self-recovered through memory refresh or inference restart. Persistent faults were mitigated through system-level resets.
These results demonstrate that COTS devices can be viable when paired with robust system-level resilience such as watchdog mechanisms, combining architectural choices, software recovery strategies, and operational mitigation. This approach enables advanced commercial silicon to support many low-Earth orbit (LEO), cislunar, and tactical defense missions at substantially lower cost than traditional rad-hard solutions.
Defense and Space Use Cases
Intelligence, surveillance, and reconnaissance (ISR) and onboard perception enable real-time object detection, tracking, and classification on unmanned aerial, terrestrial, maritime, and space platforms while reducing bandwidth requirements. Representative edge AI models include YOLOv8 or YOLOv11 for real-time detection and tracking, and MobileNet-V3 for lightweight classification and feature extraction. Other applications in development include autonomous mapping and navigation without reliance on GPS using vision-based and sensor-fusion models to support localization, obstacle avoidance, and path planning in denied or degraded environments.
Commercial and Dual-Use Applications
Earth observation satellites increasingly perform onboard scene filtering to prioritize high-value data products prior to downlink. Example edge AI models include U-Net for semantic segmentation and EfficientNet-Lite for scene classification and change detection.
Commercial drones and mobile robotics rely on edge AI for navigation, inspection, and logistics in power-constrained environments. Representative models include YOLOv8-Nano, YOLOv11-OBB for object detection and situational awareness and MiDaS for monocular depth estimation.
Edge Infrastructure AI Capabilities
Beyond mission-facing workloads, edge AI accelerators increasingly support internal infrastructure functions common to defense and commercial systems. These include power and thermal management models to dynamically manage operating states and extend component lifetime, as well as service-level optimization and operating cost reduction rely on predictive maintenance and workload optimization. Representative models include Temporal Convolutional Networks for failure prediction and compact transformer encoders such as Distil-BERT or TinyBERT, quantized for dge inference.
Additional capabilities are coming onto the scene that further enhance the power and efficiency of edge AI compute. Here are two examples.
Mixed Precision for Performance and Efficiency
Advanced edge AI accelerators increasingly support mixed-precision inference, allowing different layers of a neural network to execute using distinct numerical formats such as BF16 and INT8 within a single runtime graph. Rather than quantizing an entire model uniformly, mixed precision selectively applies higher precision to sensitivity-critical layers while maintaining lower precision elsewhere to reduce memory bandwidth and power consumption.
Transformer-based models such as DistilBERT, TinyBERT, and compact variants of Llama 3 can benefit from selective use of INT8 or BF16 precision across attention, normalization, projection, and feedforward layers. Detection models such as YOLO variants and image classification models such as EfficientNet similarly benefit from precision tuning across convolutional and head layers. In SWaP-C constrained environments, mixed precision directly improves throughput per watt and thermal efficiency while maintaining acceptable accuracy levels.
Open Vocabulary and Adaptive Recognition
Open-vocabulary inference enables models to generalize beyond fixed category sets defined at training time by leveraging shared embedding spaces, often derived from joint vision-language pretraining. This capability is increasingly valuable in defense, space, and infrastructure environments where new object classes or signal types may emerge.
Models such as OWL-ViT extend open-vocabulary capabilities to object localization. When deployed at the edge with quantized embeddings and efficient transformer backbones, open-vocabulary models reduce the need for frequent retraining or redeployment and support more adaptive ISR filtering, inspection, and autonomous response workflows.
As autonomous and distributed systems mature, the question is no longer whether AI belongs at the edge, but how to deploy it under real-world technical and business constraints. The combination of COTS edge AI accelerators, rigorous environmental testing, and system-level fault mitigation including selective generative AI workloads is reshaping design assumptions across defense, space, and commercial sectors. For engineers balancing size, weight, and power limits, performance demands, and cost, this approach offers a pragmatic and economically viable path forward.
This article was written by Stanley Crow, Vice President of Defense and Space Technology, EdgeCortix (Arlington, VA). For more information, visit here .
