story
March 3, 2026
Military autonomous systems increasingly rely on real-time perception, decision-making, and adaptation under extreme resilience constraints and increasingly smaller, heavier, and more powerful (SWaP) demands. Traditional CPU-centric embedded architectures alone cannot meet the computational demands of modern machine learning (ML) inference without exceeding the achievable power envelope or incurring unacceptable delays. Commercial off-the-shelf (COTS) edge artificial intelligence (AI) accelerators are reimagining autonomous military systems, with a particular focus on space as an operational domain. Results from recent NASA proton and heavy particle tests demonstrate that select commercial AI accelerators can exhibit radiation tolerance compatible with mission-critical autonomy, potentially changing the way defense and space programs approach onboard AI.
Autonomous military platforms, from unmanned 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 do not scale well on general-purpose embedded CPUs.
Arithmetic problems are essentially quantitative. Modern convolutional neural networks for real-time object detection can require tens of trillions of operations per second (TOPS) to maintain inference at tactically relevant frame rates. Historically, embedded CPUs capable of meeting such throughput typically exceeded acceptable power budgets or often required aggressive duty cycles that reduced responsiveness. Conversely, GPUs can provide the required computational density but can consume more than 30-40 Watts of power and are often incompatible with the size, weight, power, and cost (SWaP-C) constraints required for military platform deployment.
Latency is an equally important constraint. Autonomy requires a decisive response when communication is degraded or denied. Relying on offboard processing and cloud-based inference creates vulnerabilities to jamming, cyberattacks, and competing spectrum environments. As a result, computational bottlenecks at the edge have emerged as one of the most limiting factors for meaningful autonomy across multiple military domains.
In many autonomous military applications, delay budgets are measured in milliseconds (ms) rather than seconds. Self-driving vehicle guidance, navigation, and control loops typically require end-to-end recognition-to-action delays of less than 100 ms to maintain stability margins, with 20-50 ms delays preferred. When inference workloads exceed available onboard computing resources, the system is forced to reduce model complexity, frame rate, or autonomy level, directly constraining mission performance.
Space amplifies autonomy challenges
Space systems are the most extreme embodiment of the autonomy problem. Satellites, spacecraft, and future lunar-star platforms will need to operate with long communication delays, intermittent contact, and limited opportunities for human intervention. Onboard autonomy is not an option. Rather, it is an enabler of the mission.
From a computational perspective, space platforms combine severe size, weight, and power (SWaP) constraints with harsh radiation environments. High-energy protons and heavy ions can cause single-event effects (SEEs) that can corrupt data, disrupt functionality, or permanently damage electronic equipment. Historically, this environment has encouraged a reliance on radiation-hardened processors that trade performance and reliability, but such processors are often generations behind commercial technology and orders of magnitude higher in cost.
Radiation-hardened processors typically deliver performance in the range of single-digit to tens of giga-floating-point operations per second (GFLOPS) at power levels of 5 W to 15 W, but depending on certification level and production volume, unit costs can range from $50,000 to over $200,000 per device. This cost and performance disparity has unfortunately made high-performance onboard autonomy economically impractical for many space and defense programs, especially in a proliferating constellation.
The result was a capability gap. Advanced autonomous algorithms, especially those based on deep learning, have been difficult to deploy in space due to insufficient onboard computing power. This limitation limits applications such as autonomous rendezvous and proximity operations, onboard target recognition, adaptive payload management, and resilient spatial situational awareness.
COTS Edge AI Accelerator as a Viable Solution Path
Recent advances in commercial edge artificial intelligence (AI) accelerators represent a fundamentally different approach. Defense and space programs are increasingly evaluating commercial off-the-shelf (COTS) accelerators originally designed for ground-edge inference. These devices leverage advanced semiconductor process nodes, specialized data flows, and memory architectures optimized for neural networks to deliver orders of magnitude higher performance per watt. These improvements reduce CPU computational requirements and improve overall system performance at lower cost and power.
From a system-level perspective, this solution involves careful selection and validation rather than wholesale substitution. COTS accelerators deliver compelling benefits, including performance density measured in tens of TOPS within a single-digit watt envelope, lower unit costs with commercial volumes, and a mature software ecosystem that accelerates integrations and updates.
A remaining concern is radiation resistance. Unlike traditional space-qualified components, most COTS AI accelerators are not explicitly designed for radiation environments. This gap needs to be addressed empirically through testing and evaluation rather than assumptions.
Evidence from NASA’s proton and heavy particle tests
A recent NASA report provides quantitative insight into the issue of radiation tolerance. Under the NASA Electronic Parts and Packaging (NEPP) program, we subjected a state-of-the-art edge AI accelerator (EdgeCortix SAKURA-II) to both proton and heavy ion irradiation and characterized its susceptibility to single-event effects. Testing was performed on a Texas A&M University K500 cyclotron using linear energy transfer (LET) up to approximately 40.9 MeV·cm2/mg. (Figures 1 and 2)
This device did not demonstrate any destructive single-event latch-up events across the LET range tested. The observed effects were primarily single-event functional interrupts (SEFIs), which are often associated with PCIe interface disruptions, and single-event upsets (SEUs), which manifest as temporary changes in neural network confidence scores. Many SEUs were self-healing in subsequent inference iterations by refreshing from off-chip memory outside the illuminated 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, the fitted Weibull analysis showed an starting LET of about 0.9 MeV cm2/mg and a limiting SEFI cross section of the order of 1 × 10-4 cm2. These parameters demonstrate suitability for many missions in low-Earth orbit, geostationary orbit, and lunar-star 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.]
Impact on space-enabled AI processing
The significance of these results extends beyond a single device. Radiation tolerance is not a binary property limited to radiation-hardened components, but is actually a spectrum that can be quantified and mitigated at the system level. This answer opens up a new design space for AI accelerators, where performance per watt (and per dollar) is paramount.
From a reliability perspective, 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 abortive. System-level architectures that tolerate temporary failures through watchdog timers, task-level redundancy, memory refreshes, or periodic health checks can achieve high and efficient availability even when individual components experience recoverable events. This mindset shifts the design philosophy from eliminating all failures to limiting and managing failures, allowing the use of higher performance processors without compromising mission assurance.
By combining COTS edge AI accelerators with architectural mitigations such as checkpointing, model reloading, task-level redundancy, and autonomous recovery, space systems can achieve significantly higher performance onboard autonomy without incurring prohibitive mass or power penalties or prohibitive costs. This approach is consistent with broader defense trends toward resilient, distributed, software-defined systems.
Foundations of space and defense autonomy
Rather than being constrained by algorithms, military autonomy depends on the availability of trusted computing at the edge. Space is the most demanding test case for this challenge, combining autonomy, power limitations, and radiation exposure in a single domain.
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 toward field-adaptive, resilient autonomy at scale, commercial edge AI accelerators are increasingly well suited to play a central role.
Stanley Crow is Vice President of Defense and Space Technology at EdgeCortix. Previously, he spent 15 years at Northrop Grumman in positions including CTO and Vice President of Technology, Engineering and Manufacturing for the Defense Systems Division and Chief Executive Officer of Northrop Grumman Japan. Prior to joining Northrop Grumman, Mr. Crow was an Associate Principal at McKinsey & Company, serving technology, aerospace and defense clients around the world. In addition, he served nearly 30 years in the U.S. Air Force, including both active duty and reserves focused on space, intelligence, and advanced capabilities needs in the Indo-Pacific region.
EdgeCortix https://www.edgecortix.com/en/
