Space-grade ML inference requires knowing exactly how your accelerator fails under radiation. Commercial, closed-source accelerators make that impossible: you can observe failure outputs, but you cannot instrument the internal state, reroute fault propagation, or design targeted mitigation at the register-transfer level. The Tensil NN accelerator -- RTL-accessible, deployed on a Zynq UltraScale+ -- fixes this. This paper delivers the foundational irradiation dataset: 4.29 x 10^10 protons/cm^2 at 20-58 MeV, seven workload interruptions, two output-corruption events returning wrong CIFAR-10 classes, with full kernel logs and restart-sequence characterization.
The contribution is not the failure numbers themselves. It is that these numbers are now actionable. Because the fault injection team has RTL access, they can trace the two corruption events to specific flip-flop domains, design selective TMR or scrubbing strategies, and re-run the characterization to verify mitigation effectiveness. With a closed commercial IP block, the same team would be writing a datasheet footnote about "radiation sensitivity TBD" and waiting for the vendor to release a rad-hard variant on their own schedule.
This is the open-source hardware argument made concrete in a domain where correctness is safety-critical. COTS FPGA inference is already used in small satellites, UAVs, and experimental aerospace platforms where rad-hard ASICs are cost-prohibitive. The Tensil characterization baseline is the first step toward a repeatable, community-maintained fault model for open-source ML accelerators -- the equivalent of what decades of SRAM SEU characterization data did for space-qualified memory design.
Aerospace primes and smallsat bus vendors that have been defaulting to rad-hard processor families for ML inference should track this. In 2-3 years, if the community accumulates baselines across multiple open-source accelerator architectures (Tensil, VTA, NVDLA subsets), the gap between "closed-source radiation assurance" and "open-source radiation characterization" closes. The open-source path will not be cheaper because the hardware is free. It will be cheaper because the fault model is auditable.