Where int8 breaks
Every chip we support runs fastest in 8-bit integers. Converting a model from 16-bit floats to int8 usually costs less than a point of accuracy on a benchmark, and buys a two to four times speedup. That trade is so reliable that most teams stop measuring it. We kept measuring, and found that the loss is not spread evenly.
In well-lit regions the int8 model matches the float model almost exactly. In the darkest tenth of a frame, recall drops by up to 19 points. The cause is the calibration step. Ranges are chosen from typical images, which are mostly bright, so the small activations that carry information in shadows all collapse into the same few integer values.
Ranges that follow the light
Our method calibrates ranges per channel and per exposure band. During compilation, calibration frames are split by scene brightness and each band gets its own statistics. At run time, the Runtime reads the exposure the image processor already reports and picks the matching set of scales. The switch costs no extra compute, because the scales fold into the first convolution.
parallax compile depth-s.pxm \ --target hexagon \ --precision int8 \ --calibrate exposure-bands=4 \ --frames ./calib/dock-03
Four bands are enough for every sensor we have tested. More bands improve the darkest regions slightly, but the gain stops being measurable beyond six.
Co-designing the architecture
Quantization is not only a compiler problem. Some layers are fragile in int8 no matter how ranges are chosen, typically wide activations right after a normalisation. We search architectures with the quantized model in the loop, so those layers are replaced before training rather than patched afterwards.
Measure int8 where the camera struggles, not where the benchmark is easy.
Results
On our low-light evaluation set, exposure-banded int8 Depth-S keeps 98.7 percent of float recall in the darkest decile, compared with 81 percent for standard per-tensor calibration. Latency is unchanged at 6.4 ms on Hexagon. The method is on by default in Runtime 3.3 and later.