COTS edge AI accelerators – an enabler for military autonomous systems in all domains, including space

Story

March 03, 2026

Autonomous military platforms – ranging from uncrewed aerial vehicles and autonomous ground systems to maritime and space assets – are increasingly expected to operate with minimal human oversight. These systems rely on deep neural networks for tasks such as object detection, sensor fusion, navigation, and adaptive mission planning. The computational profile of these workloads is dominated by matrix-intensive operations that scale poorly on general-purpose embedded CPUs.

The operation problem is fundamentally quantitative: A modern convolutional neural network for real-time object detection can require tens of trillions of operations per second (TOPS) to sustain inference at tactically relevant frame rates. Historically, embedded CPUs capable of meeting such throughput typically exceeded acceptable power budgets or often required aggressive duty cycling that would degrade responsiveness. Conversely, GPUs can provide the necessary compute density, but can consume power in excess of 30 to 40 watts and are often incompatible with size, weight, power, and cost (SWaP-C) constraints required in military platform deployments.

Latency is an equally critical constraint. Autonomy demands deterministic response under degraded or denied communications. Reliance on off-board processing or cloud-based inference introduces vulnerability to jamming, cyberattack, and contested spectrum environments. As a result, the computational bottleneck at the edge has emerged as one of the most limiting factors for meaningful autonomy across multiple military domains.

In many autonomous military applications, latency budgets are measured in milliseconds (ms) rather than seconds. Guidance, navigation, and control loops for autonomous vehicles typically require end-to-end perception-to-action latencies below 100 ms to maintain stability margins, with latencies of 20 ms to 50 ms desired. When inference workloads exceed available onboard compute resources systems are forced to reduce model complexity, frame rate, or autonomy level – directly constraining mission performance.

Space amplifies the autonomy challenge

Space systems represent the most extreme instantiation of the autonomy problem. Satellites, spacecraft, and future cislunar platforms must operate with long communication delays, intermittent contact, and limited opportunities for human intervention. Onboard autonomy is not optional; rather, it is mission-enabling.

From a computational perspective, space platforms combine severe size, weight, and power (SWaP) constraints with an unforgiving radiation environment. High-energy protons and heavy ions can induce single-event effects (SEEs) that corrupt data, interrupt functionality, or permanently damage electronics. Historically, this environment drove reliance on radiation-hardened processors that trade performance for reliability, but such processors often lagged commercial technology by multiple generations and cost orders of magnitude more.

Radiation-hardened processors typically deliver performance in the range of single-digit to low-tens of giga floating-point operations per second (GFLOPS) at power levels of 5 W to 15 W, while unit costs can exceed upwards of $50,000 to $200,000 per device depending on qualification level and production volume. This cost and performance disparity has unfortunately made high-performance onboard autonomy economically impractical for many space and defense programs, particularly for proliferated constellations.

The result has been a capability gap. Advanced autonomy algorithms – particularly those based on deep learning – have been difficult to deploy in space due to insufficient onboard compute power. This limitation constrains applications such as autonomous rendezvous and proximity operations, onboard target recognition, adaptive payload management, and resilient space situational awareness.

COTS edge AI accelerators as a viable solution path

Recent advancements in commercial edge artificial intelligence (AI) accelerators point to a fundamentally different approach. Defense and space programs are increasingly evaluating commercial off-the-shelf (COTS) accelerators originally designed for terrestrial edge inference. These devices leverage advanced semiconductor process nodes, specialized data flows, and memory architectures optimized for neural networks, achieving orders-of-magnitude improvements in performance per watt. These improvements reduce compute requirements on the CPU, enabling greater overall system performance at lower cost and power.

From a system-level perspective, the solution involves careful selection and validation rather than wholesale substitution. COTS accelerators enable compelling advantages, such as performance density measured in tens of TOPS within single-digit watt envelopes, reduced unit cost through commercial volumes, and mature software ecosystems that accelerate integration and updates.

A remaining concern has been radiation resilience. Unlike traditional space-qualified parts, most COTS AI accelerators are not designed explicitly for radiation environments. This gap must be addressed empirically through test and evaluation rather than assumption.

Evidence from NASA proton and heavy-ion testing

A recent NASA report provides quantitative insight into the radiation-resilience question. Under the NASA Electronic Parts and Packaging (NEPP) Program, a modern edge AI accelerator (the EdgeCortix SAKURA-II) was subjected to both proton and heavy-ion irradiation to characterize its susceptibility to single-event effects. Testing was conducted at the Texas A&M University K500 cyclotron using linear energy transfers (LETs) up to approximately 40.9 MeV·cm²/mg. (Figures 1 & 2.)

The device demonstrated no destructive single-event latchup events across the tested LET range. Observed effects were primarily single-event functional interrupts (SEFIs), often associated with PCIe interface disruptions, and single-event upsets (SEUs) manifesting as transient changes in neural network confidence scores. Many SEUs were self-recovering on subsequent inference iterations due to refresh from off-chip memory outside the irradiation zone.

[Figure 1 and Figure 2 ǀ An edge AI accelerator, subjected to irradiation to characterize its susceptibility to single-event effects, demonstrated no destructive single-event latchup events across the tested linear energy transfer (LET) range. Source: NASA.]

Quantitatively, fitted Weibull analysis indicated an onset LET of approximately 0.9 MeV·cm²/mg and a limiting SEFI cross-section on the order of 1×10-4 cm². These parameters indicate suitability for many missions in low Earth orbit, geosynchronous orbit, and cislunar regimes. (Figure 3.)

[Figure 3 ǀ Fitted Weibull analysis indicated an onset LET of approximately 0.9 MeV·cm²/mg and a limiting SEFI cross-section on the order of 1×10-4 cm². These parameters indicate suitability for many missions in low Earth orbit, geosynchronous orbit, and cislunar regimes. Source: NASA.]

Implications for space-ready AI processing

The significance of these results extends beyond a single device. Radiation tolerance is not a binary property limited to radiation-hardened parts, but is in fact a spectrum that can be quantified and mitigated at the system level. For AI accelerators, where performance per watt (and per dollar) is paramount, this answer opens a new design space.

From a reliability standpoint, typical single-event upset rates for advanced microelectronics in low Earth orbit are often on the order of 10-6 to 10-4 upsets per device-day, depending on shielding, orbit, and solar conditions. Importantly, not all upsets are mission-terminating. System-level architectures that tolerate transient faults – through watchdog timers, task-level redundancy, memory refresh, or periodic health checks – can achieve high effective availability even when individual components experience recoverable events. This mindset shifts the design philosophy from eliminating all faults to bounding and managing them, enabling the use of higher-performance processors without compromising mission assurance.

By combining COTS edge AI accelerators with architectural mitigations, such as checkpointing, model reloads, redundancy at the task level, and autonomous recovery, space systems can field substantially more capable onboard autonomy without prohibitive mass or power penalties or exorbitant costs. This approach aligns with broader defense trends toward resilient, distributed, and software-defined systems.

A foundation for space and defense autonomy

Military autonomy is no longer constrained by algorithms, but by the availability of reliable compute at the edge. Space represents the most demanding test case for this challenge, combining autonomy, power limits, and radiation exposure in a single domain.

The emerging evidence suggests that carefully selected and validated COTS edge AI accelerators can form the computational backbone of next-generation autonomous systems, including space platforms. As defense programs move to field adaptive, resilient autonomy at scale, commercial edge AI accelerators are well suited to play an increasingly central role.

Stanley Crow is EdgeCortix’s Vice President of Defense & Space Technology. He previously served for 15 years in Northrop Grumman in roles including CTO and VP Technology, Engineering and Manufacturing for the company’s Defense Systems sector and Chief Executive for Northrop Grumman Japan. Prior to joining Northrop Grumman, Mr. Crow was an Associate Principal at McKinsey and Company, serving technology, aerospace, and defense clients globally. Additionally, he served almost three decades in the U.S. Air Force, including both active-duty and reserve assignments focused on space, intelligence, and advanced capability needs in the IndoPacific.

EdgeCortix    https://www.edgecortix.com/en/