Selecting edge AI hardware for a defense system is not a performance-only decision. A platform that delivers the highest TOPS count may be disqualified by its power consumption, its operating temperature range, or its export classification. Hardware that easily passes MIL-STD qualification for one platform class — ground vehicle — may be entirely unsuitable for a UAV payload where every gram and milliwatt counts. This article provides a structured comparison of the three dominant edge AI accelerator families for defense applications, with a decision framework for common platform types.

Selection Criteria: TOPS, TDP, Operating Temperature, MIL-STD Compliance

TOPS (Tera Operations Per Second) measures the theoretical peak throughput for 8-bit integer operations, which is the primary precision used in deployed inference. TOPS numbers are useful for comparison within a product family but misleading across families because different architectures achieve different real-world model throughputs for the same TOPS rating. Always benchmark with the specific model and precision you intend to deploy, not the manufacturer's TOPS headline.

TDP (Thermal Design Power) defines the maximum sustained power draw the hardware is designed to dissipate. For battery-powered platforms, TDP is a hard budget constraint. A dismounted soldier's battery pack may support 5–10W of dedicated compute; a small UAV payload budget might be 3–8W; a vehicle-mounted system might allow 50–200W. The TDP of the module is only part of the story — the carrier board, power regulation circuits, memory, and peripherals add additional consumption.

Operating temperature for commercial hardware is typically 0°C to +70°C. Defense-grade requirements typically demand −40°C to +85°C for ground vehicle applications and −54°C to +85°C for aviation applications. Exceeding the operating temperature range causes thermal throttling, reduced performance, and eventually hardware failure. Modules that are not rated for the full military temperature range require additional thermal engineering — heaters for cold environments, enhanced cooling for high-temperature environments — that adds mass, volume, and complexity.

MIL-STD compliance covers shock resistance (MIL-STD-810H Method 516.8), vibration (Method 514.8), humidity (Method 507.6), altitude (Method 500.6), and EMI/EMC (MIL-STD-461). Consumer and industrial modules are typically not MIL-STD qualified; the system integrator must qualify the complete hardware assembly including enclosure and mounting. Some vendors — particularly those targeting defense — offer pre-qualified module assemblies that simplify this process.

NVIDIA Jetson Orin Family: AGX Orin, Orin NX, Orin Nano

The Jetson Orin family spans three product tiers sharing the same Ampere GPU architecture and Arm Cortex-A78AE CPU cores, differing in GPU core count, CPU core count, memory capacity, and power range:

Jetson AGX Orin (32GB and 64GB variants) — 275 TOPS at up to 60W, or 67 TOPS in 15W mode. 12 Ampere GPU SM cores, 12 Cortex-A78AE CPU cores, 64 GB LPDDR5 (on the 64GB variant). The highest performance option in the Jetson lineup. Pre-qualified carrier boards with extended temperature operation are available from several ruggedized computer vendors. The primary platform for vehicle-mounted and ground station AI systems where power budget allows.

Jetson Orin NX — available in 8GB and 16GB variants at 70–100 TOPS (10–25W). 8 Ampere GPU SM cores in the 16GB variant. A strong mid-range option for systems where the AGX Orin's power and physical size are excessive but significant inference capability is still needed. Frequently used in small-form-factor UAV ground control stations and dismounted computing platforms.

Jetson Orin Nano — 20–40 TOPS at 7–15W. The entry-level Orin family member, targeting applications where the prior-generation Jetson TX2 or Nano was used. 4 GB or 8 GB LPDDR5. Suitable for sensor preprocessing and lightweight inference on power-constrained platforms but cannot run large vision models at real-time frame rates.

All Orin family modules use NVIDIA's CUDA and TensorRT ecosystem, enabling straightforward model optimization from PyTorch training to Jetson deployment. DeepStream SDK provides optimized multi-camera pipeline support. The Orin GPU architecture includes dedicated tensor cores for INT8 inference, delivering roughly 4× the throughput of FP32 for vision models. JetPack SDK provides a unified software stack with CUDA, cuDNN, TensorRT, VPI, and multimedia APIs pre-installed.

Hailo-8 and Hailo-8L: Low-Power AI Acceleration

Hailo's Dataflow Architecture differs fundamentally from GPU-based inference. Rather than a general-purpose parallel compute array, Hailo uses a custom dataflow graph compiler that maps CNN models directly onto a network of specialized compute elements, eliminating the data movement overhead that dominates GPU energy consumption for inference workloads. This architecture delivers high TOPS per watt for CNN inference specifically, at the cost of reduced flexibility for model architectures that don't map efficiently to the Hailo graph.

The Hailo-8 delivers 26 TOPS at approximately 2.5W peak power in a 15×15mm BGA package or PCIe M.2 card. The operating temperature range of the commercial module is −40°C to +85°C — unusual for a commercial device and qualifying it for direct use in many defense applications without thermal engineering overhead. A YOLOv5s model compiled for Hailo-8 runs at approximately 120 fps, and YOLOv8s at approximately 60 fps, at approximately 1.5W active power — a TOPS-per-watt efficiency advantage of roughly 3–5× over comparable Jetson hardware for these specific workloads.

The Hailo-8L is a lower-power variant delivering 13 TOPS at approximately 1W, targeting wearable and IoT edge applications. For UAV payloads where the total compute power budget is under 3W, the Hailo-8L is frequently the only option that meets both performance and power requirements.

The key ecosystem constraint: the Hailo Dataflow Compiler requires model conversion from ONNX format, and not all layer types are natively supported. Models with custom layers, attention mechanisms, or transformer architectures may require adaptation. The Hailo Model Zoo provides pre-optimized versions of common defense-relevant models (YOLOv5, YOLOv8, ResNet) that can be used directly without compiler expertise.

Intel Movidius Myriad X and OpenVINO

The Myriad X VPU (Vision Processing Unit) integrates 16 SHAVE vector processors with a dedicated Neural Compute Engine (NCE) delivering approximately 4 TOPS at ~1W average power. In its PCIe form factor (Intel Neural Compute Stick 2 or M.2 module) it provides a convenient AI acceleration add-on for systems already running on Intel x86 or Atom processors.

Intel's OpenVINO toolkit is the primary differentiator. OpenVINO provides a model optimization and deployment pipeline that supports heterogeneous execution across Intel CPU, iGPU, VPU, and FPGA targets with a single API. A model deployed through OpenVINO can be executed on whichever Intel hardware is available without code changes — useful for programs where the compute platform may vary across hardware generations. OpenVINO's model optimizer supports import from TensorFlow, PyTorch (via ONNX), Caffe, and PaddlePaddle.

The primary defense application for Myriad X is embedded vision preprocessing in systems with tight power envelopes and existing Intel ecosystem dependencies — Intel RealSense depth cameras, Intel Atom-based embedded computers. For standalone high-performance inference, the Hailo-8 or Jetson family are typically preferred.

Key insight: Don't choose edge AI hardware based on peak TOPS alone. Run your target model at your target precision (INT8, FP16) on each platform and measure actual latency and throughput. A 275-TOPS Jetson AGX Orin running a YOLOv8-nano at 200fps does not make the hardware choice better if your constraint is 3W power budget — a 26-TOPS Hailo-8 at 1.5W running the same model at 60fps may be the correct answer.

Decision Framework: UAV Payload vs Vehicle-Mounted vs Wearable

Small UAV payload (under 500g, 3–8W compute budget): Hailo-8 or Hailo-8L. The combination of low power, low weight (the Hailo-8 M.2 card weighs approximately 6g), and adequate inference performance for vision-based detection makes this the dominant choice. Pair with a lightweight Arm SoC (Raspberry Pi CM4, NXP i.MX 8) for system control and communications.

Medium UAV or tactical vehicle payload (5–50W compute budget): Jetson Orin NX (16GB). Sufficient performance for YOLOv8-large, multi-sensor fusion, and concurrent tracking algorithms, within a 15–25W power envelope. Available in ruggedized carrier board configurations from vendors like Connect Tech and ADLINK.

Vehicle-mounted or ground station system (50–200W compute budget): Jetson AGX Orin (64GB). Full TensorRT ecosystem, DeepStream multi-camera support, concurrent AI workload support, and 64GB unified memory for large models including LLM inference for ISR triage assistance.

Wearable / dismounted soldier system (under 5W compute budget): Hailo-8L (1W) or Movidius Myriad X (1W) paired with ultra-low-power Arm SoC. Performance limited to lightweight detection and classification models; complex tracking algorithms require offloading to vehicle or base station systems.