Session
Advanced Technologies 2
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
The deployment of onboard artificial intelligence (AI) on small satellites enables real-time Earth-observation triage, object detection, and event-driven tasking with reduced dependence on downlink. These capabilities rely on the integrity of neural network parameters (weights and biases) stored across both volatile memory and non-volatile storage (e.g., DRAM/SRAM and SSD/SD/UFS/eMMC). In low-Earth orbit (LEO), radiation-induced upsets can flip bits in commercial-off-the-shelf (COTS) memory and storage, corrupting model parameters and degrading inference accuracy. Such degradation can manifest as persistent misclassification or missed detections. Existing mitigation options are often difficult to scale within small-satellite constraints: Triple Modular Redundancy (TMR) incurs substantial mass and power overhead, while common coding-based EDAC such as SECDED provides limited protection as multi-bit errors become more likely in high-density devices.
This paper presents ZEDAC, a multi-bit EDAC architecture that protects AI model data both at rest (in storage) and during use (in memory). ZEDAC generates protection metadata when model data is written and performs integrity checking and correction when the data is later read, enabling restoration of corrupted model parameters without retraining. ZEDAC is deployed with a radiation-tolerant companion EDAC module accessed over SPI. The module stores protection metadata in TMR FRAM and employs a rad-hard voter to reduce the likelihood of metadata corruption, while the primary model data remains in COTS memory and storage. We validated ZEDAC across heterogeneous compute platforms, including PCs, NVIDIA Jetson Orin Nano, ZSOM-M01 (radiation-tolerant microcontroller-based system-on-module), and Raspberry Pi 5 with an AI HAT, on workloads spanning CNN, YOLOv8, and quantized large-language-model inference.
ZEDAC applies extended Hamming codes along both the row and column dimensions of each protected data block, enabling structured multi-bit correction with controlled overhead. The configuration used in this work corrects up to three erroneous bits per row, and can correct larger numbers of total bit flips within a block when errors are distributed across rows and columns; correction effectiveness depends on the observed error distribution and burst characteristics. We evaluate ZEDAC end-to-end using representative AI workloads (CNN classification and YOLO-class object detection) under both random and burst corruption regimes applied to memory-resident parameters and stored model files. We combine software-based fault injection (to sweep error rates, burst lengths, and spatial distributions) with hardware-based fault injection on COTS DRAM, using controlled optical stimulation on an exposed die to induce upset-like bit flips and corroborate behavior on physical memory hardware. Without protection, both models exhibit severe accuracy degradation under injected faults. With ZEDAC enabled, model performance can be restored toward near-baseline levels within minutes when the corruption patterns fall within the correctable regimes, without retraining or ground intervention. We characterize correction success and runtime overhead across a range of fault regimes, and show that recovery effectiveness varies with the observed corruption pattern (random versus burst, density, and locality). Overall, ZEDAC provides a practical approach for improving the robustness of onboard AI on small satellites by protecting model integrity across the full lifecycle from storage to memory to inference using COTS devices augmented with a compact radiation-tolerant protection module.
Document Type
Event
Radiation-Resilient AI for Small Satellites: A Universal Multi-Bit EDAC Architecture for Protecting On-Board Neural Network Models in Memory and Storage
Salt Palace Convention Center, Salt Lake City, UT
The deployment of onboard artificial intelligence (AI) on small satellites enables real-time Earth-observation triage, object detection, and event-driven tasking with reduced dependence on downlink. These capabilities rely on the integrity of neural network parameters (weights and biases) stored across both volatile memory and non-volatile storage (e.g., DRAM/SRAM and SSD/SD/UFS/eMMC). In low-Earth orbit (LEO), radiation-induced upsets can flip bits in commercial-off-the-shelf (COTS) memory and storage, corrupting model parameters and degrading inference accuracy. Such degradation can manifest as persistent misclassification or missed detections. Existing mitigation options are often difficult to scale within small-satellite constraints: Triple Modular Redundancy (TMR) incurs substantial mass and power overhead, while common coding-based EDAC such as SECDED provides limited protection as multi-bit errors become more likely in high-density devices.
This paper presents ZEDAC, a multi-bit EDAC architecture that protects AI model data both at rest (in storage) and during use (in memory). ZEDAC generates protection metadata when model data is written and performs integrity checking and correction when the data is later read, enabling restoration of corrupted model parameters without retraining. ZEDAC is deployed with a radiation-tolerant companion EDAC module accessed over SPI. The module stores protection metadata in TMR FRAM and employs a rad-hard voter to reduce the likelihood of metadata corruption, while the primary model data remains in COTS memory and storage. We validated ZEDAC across heterogeneous compute platforms, including PCs, NVIDIA Jetson Orin Nano, ZSOM-M01 (radiation-tolerant microcontroller-based system-on-module), and Raspberry Pi 5 with an AI HAT, on workloads spanning CNN, YOLOv8, and quantized large-language-model inference.
ZEDAC applies extended Hamming codes along both the row and column dimensions of each protected data block, enabling structured multi-bit correction with controlled overhead. The configuration used in this work corrects up to three erroneous bits per row, and can correct larger numbers of total bit flips within a block when errors are distributed across rows and columns; correction effectiveness depends on the observed error distribution and burst characteristics. We evaluate ZEDAC end-to-end using representative AI workloads (CNN classification and YOLO-class object detection) under both random and burst corruption regimes applied to memory-resident parameters and stored model files. We combine software-based fault injection (to sweep error rates, burst lengths, and spatial distributions) with hardware-based fault injection on COTS DRAM, using controlled optical stimulation on an exposed die to induce upset-like bit flips and corroborate behavior on physical memory hardware. Without protection, both models exhibit severe accuracy degradation under injected faults. With ZEDAC enabled, model performance can be restored toward near-baseline levels within minutes when the corruption patterns fall within the correctable regimes, without retraining or ground intervention. We characterize correction success and runtime overhead across a range of fault regimes, and show that recovery effectiveness varies with the observed corruption pattern (random versus burst, density, and locality). Overall, ZEDAC provides a practical approach for improving the robustness of onboard AI on small satellites by protecting model integrity across the full lifecycle from storage to memory to inference using COTS devices augmented with a compact radiation-tolerant protection module.
