Why can an edge model fit in flash but still run out of RAM?

Weights are only one part of the memory picture. Intermediate activations, input windows, preprocessing buffers and operator scratch space can require RAM. The firmware also needs stacks, communication buffers and application state. Inspect the generated deployment estimates and linker map, but verify peak runtime use on the target because static allocation alone does not describe every execution path.

For a simple example, a 128 × 128 RGB image uses 49,152 bytes at one byte per channel before preprocessing or model activations. Keeping two such frames roughly doubles that input-buffer requirement. Changing the input size or buffer lifetime may save more memory than a small reduction in model weights. Check whether buffers coexist before adding all nominal sizes or assuming they can overlap.

Exercise the worst relevant combination of inference, networking and logging. Monitor stack margins and handle allocation failures explicitly. If memory is tight, consider streaming preprocessing, shorter windows or a different architecture, then retest accuracy and timing. A deployment estimate is useful for selection, while the integrated firmware test determines whether the actual application has adequate margin.