Start at the image source
The original Raspberry Pi AI HAT+ arrived in October 2024 in 13 and 26 TOPS variants. Raspberry Pi described integration with its existing Hailo camera software path. For a vision project, start the design one step earlier than model execution: what image reaches the application, at what resolution, and how old is it? A fast detector can still produce an unhelpful result if it works through a queue of stale images from the capture stage.
Name the work between frames
Draw capture, format conversion, resizing, inference, post-processing and delivery as separate stages. For each boundary, write the data format and who owns the buffer. This is a design exercise, not a claim that every implementation performs a copy. It exposes where a proposed change needs verification. Replacing one stage with a faster implementation may simply move the waiting elsewhere, particularly if output handling or display work runs in the same application thread.
Decide what overload means
A camera normally keeps producing frames when inference falls behind. Choose whether the application must process every frame or should discard old work to favour recent observations. An archival inspection task and a live steering aid can justify different policies. Give the queue a limit and document the action at that limit. Otherwise a temporary slowdown can quietly become an increasing delay, while an inference-only rate continues to look acceptable in a dashboard.
Compare the same complete task
Raspberry Pi notes that models compiled for Hailo-8L can run on Hailo-8, while the reverse is not guaranteed. Preserve the exact board variant and compiled model in the project record. When considering a change, use the same source images and output requirement on both configurations. Report capture-to-result delay and dropped work beside accelerator timing. The decision then concerns a camera application with a defined response policy, rather than a ranking of two published compute figures.
Sources & publication notes
Retrospective period: February 2025. Written and published November 2, 2026. Product announcements establish release timing; support pages are linked as current references.