G · Hardware tiers and conclusions
What the measurements say about constrained hardware
Tier A · measured
Development laptop
Full machine, 4 cores, 1 core (CPU affinity and thread limits; no memory cap, no clock cap).
Scaling with core count on the development laptop. One-way inference only: a miss here makes a miss on a Raspberry Pi 5 very likely; a pass says nothing about the Pi.
Tier B · pending first CI run
ARM cloud VM
Real ARM server-class shared cloud VM (GitHub Actions). Noisy. NOT a Raspberry Pi.
The workflow is written and validated locally but has not run: it needs a push.
Tier C · UNAVAILABLE
Raspberry Pi 5
Not measured. Run the same CLI with --profile pi5 on a real Pi 5; the CPU model in run.json identifies the hardware.
What can be inferred
- newer_college_bag9 · baseline
laptop-full vs laptop-4c: per-repeat median ranges overlap (674-839 ms vs 649-739 ms); no difference is claimed. - newer_college_bag9 · baseline
laptop-4c vs laptop-1c: per-repeat median ranges overlap (649-739 ms vs 650-676 ms); no difference is claimed. - newer_college_bag9 · baseline
Profiles were measured at different times on a shared, non-isolated laptop; a non-monotonic ordering across profiles reflects machine drift as much as core count. - newer_college_bag9 · baseline
[ASSUMED, one-way inference, not measured] It misses the deadline on 4 laptop cores, so a Raspberry Pi 5 very likely misses it too. - oxford_spires · baseline
laptop-4c is faster than laptop-full beyond run-to-run spread (median 438 ms vs 522 ms). - oxford_spires · baseline
laptop-4c is faster than laptop-1c beyond run-to-run spread (median 438 ms vs 779 ms). - oxford_spires · baseline
Profiles were measured at different times on a shared, non-isolated laptop; a non-monotonic ordering across profiles reflects machine drift as much as core count. - oxford_spires · baseline
[ASSUMED, one-way inference, not measured] It misses the deadline on 4 laptop cores, so a Raspberry Pi 5 very likely misses it too. - newer_college_bag9 · optimized
laptop-full is faster than laptop-4c beyond run-to-run spread (median 324 ms vs 611 ms). - newer_college_bag9 · optimized
laptop-4c is faster than laptop-1c beyond run-to-run spread (median 611 ms vs 870 ms). - newer_college_bag9 · optimized
Profiles were measured at different times on a shared, non-isolated laptop; a non-monotonic ordering across profiles reflects machine drift as much as core count. - newer_college_bag9 · optimized
[ASSUMED, one-way inference, not measured] It misses the deadline on 4 laptop cores, so a Raspberry Pi 5 very likely misses it too.
Established
- Latency, tail latency and deadline-violation rate of this implementation on this laptop, per profile.
- That limiting cores changes (or does not change) the result.
Not established
- Raspberry Pi 5 or any embedded performance. A miss on 4 laptop cores makes a Pi miss very likely; a pass says nothing about the Pi.
- Memory-bandwidth, cache and thermal behaviour of embedded boards.
- Hard real-time guarantees from Python.
- Flight conditions: vibration, IMU fusion, wind.
Jetson and GPU execution are out of scope: the pipeline is CPU-only. A GPU path would be a separate project.