This companion page maps the actual shape of the literature: what tasks dominate, where FPGA designs fit in the spacecraft pipeline, and where “space-ready” claims thin out once radiation and fault tolerance enter the picture.
Spacecraft do not get to offload everything to Earth. This flow shows why sensor workloads, bandwidth pressure, and autonomy demands all converge on local inference hardware.
Highlighted Stage
Constraint Load
Downlink scarcity0.93
Latency to ground0.89
Power ceiling0.84
Thermal rejection0.78
Mission risk asymmetry0.95
Task mix seen in the episode
VisionAutonomyCompression / selection
Most cited examples still sit in Earth-observation imaging, even when the framing expands to “space AI.”
Time-Multiplexed vs Dataflow FPGA Design
The survey reports a real architectural split. Toggle the mode to see how buffering, reuse, latency, and memory traffic trade against each other.
Architecture readout
Layer scheduling heatmap
Idle / low useModerateHot / bottleneck
Evidence Map: Performance Claims vs Mission Readiness
Mock points summarize the literature pattern described in the episode: many strong speed or efficiency demonstrations, fewer studies that move into radiation and dependability evidence.
Interpretation
Corpus slices
CNN-focused studies34
Segmentation / detection emphasis17
Clear radiation or fault eval9
Transformers or language-like~0
Commercial-board dominantHigh
Reliability Is a System Property, Not a MAC Property
Faults do not stay politely inside multiply-accumulate units. Hover the subsystem matrix and compare protection strategies to see where the survey’s caution lands.
Low fault impactMediumHigh fault impact
Takeaway
References
Primary survey, reliability foils, resilient space deployments, and prior AI Post Transformers episodes referenced in the transcript.
1
FPGA-Based Neural Network Accelerators for Space Applications: A Survey — Pedro Antunes, Artur Podobas, 2025
An FPGA-Based Hardware Accelerator for CNNs Inference on Board Satellites: Benchmarking with Myriad 2-Based Solution for the CloudScout Case Study — Rapuano et al., 2021