SpaceExit: Enabling Efficient Adaptive Computing in Space with Early Exits
summary
System builds an onboard pipeline that tiles images, runs a multiscale feature backbone with early exits, and schedules tiles across heterogeneous accelerators while managing power and thermal state with DVFS. Evaluation on a standard object‑detection dataset across multiple hardware setups and link conditions shows normalized useful-throughput gains of 5.2% to 37.6% over four baselines and a slight MAP improvement, demonstrating the tile-level scheduler plus resource controller actually increases delivered insights from orbit without losing accuracy.
AI-written summary · about
related talks
22mPreview: Hardware and Offloads
06mMETASAT Demonstrator: Mixed Criticality, Accelerated AI Computing for Future Space Systems
36mAccelerating AI training fleets with sched_ext
19mOffloading-Efficient Sparse AI Systems
56mAutomatically Scalable Computation
15mUSHER: Holistic Interference Avoidance for Resource Optimized ML Inference
15mECHO: Efficient KV Cache Offloading with Lossless Prefetching for Serving Native...
45mHigh-Performance and Scalable Middleware for HPC and Deep...