SI & computation
Choose the right task, model, hardware and moment—then measure the complete result.
The intelligence layer
SI is the intelligence theme in SpaceXSI. The practical research here concerns learning-based systems, inference and resource optimization. Original source titles retain the terminology used by their publishers.
The useful question is which system can satisfy a task reliably within an energy and latency budget. Some tasks suit small specialized models; others require larger systems. Conventional software remains appropriate where a model contributes little value.
Training and inference
Training develops model parameters from data. Inference applies a trained model to new inputs. Their execution patterns and constraints differ, so they need separate evaluations.
Measure the full request lifecycle: input preparation, execution, memory movement, postprocessing and retries. Token count or accelerator activity alone is not a complete definition of useful output.
An optimization stack
NVIDIA’s TensorRT documentation describes lower-precision inference formats and quantization schemes. Whether an implementation saves energy must be tested on its hardware and workload; precision alone supplies no universal savings percentage.
| Technique | Opportunity | Trade-off to test |
|---|---|---|
| Task routing | Use a sufficient smaller system | Routing mistakes and quality regressions. |
| Quantization | Reduce numerical precision and storage | Accuracy, hardware support and overhead. |
| Batching | Share execution resources | Queue time and latency. |
| Caching | Reuse valid results | Staleness, privacy and hit rate. |
| Flexible scheduling | Move work within deadlines | Data transfer and delayed completion. |
Space-native intelligence
NASA’s EO-1 mission used the Autonomous Science Experiment for onboard intelligent scheduling. This is a historical example of useful autonomy in orbit.
A research opportunity is to extract relevant information near the sensor before downlink. Our inference is that compact results may reduce communication requirements, but that benefit must be weighed against onboard power, compute mass and reliability.
A proposed evaluation sequence
- Publish a fixed task suite and quality criteria.
- Change one optimization at a time against a documented baseline.
- Measure energy, latency, accuracy and retries together.
- Include warm-up and idle conditions in the report.
- Repeat under representative loads rather than only full utilization.