Session

Poster Session 2

Location

Salt Palace Convention Center, Salt Lake City, UT

Abstract

MOTIVATION:

SmallSats are asked to make time-sensitive calls (vessel detection, cloud filtering, debris tracking) but CubeSat-class compute is starved for power, memory, and radiation-tolerant silicon.

Most missions still downlink raw imagery for ground processing, costing hours-to-days of latency and scarce bandwidth. Running inference onboard means solving two separate problems:

  • When to run inference: compute and power budgets can’t tolerate a neural network firing on every frame.
  • Whether it fits when it does run: the model has to live inside a memory and compute envelope measured in megabytes, not gigabytes.

Document Type

Event

SSC26-P2-66 (2).pdf (469 kB)
Paper

Available for download on Saturday, August 22, 2026

Share

COinS
 
Aug 24th, 12:00 AM

Advanced AI Deployment on SmallSats: A Mixed Precision Approach to Onboard Perception and FPGA-Gated Inference

Salt Palace Convention Center, Salt Lake City, UT

MOTIVATION:

SmallSats are asked to make time-sensitive calls (vessel detection, cloud filtering, debris tracking) but CubeSat-class compute is starved for power, memory, and radiation-tolerant silicon.

Most missions still downlink raw imagery for ground processing, costing hours-to-days of latency and scarce bandwidth. Running inference onboard means solving two separate problems:

  • When to run inference: compute and power budgets can’t tolerate a neural network firing on every frame.
  • Whether it fits when it does run: the model has to live inside a memory and compute envelope measured in megabytes, not gigabytes.