The simulation visualizes a layer of floating-point weights being mapped onto a discrete integer grid, showing how scale and zero-point compress the continuous distribution into quantization buckets.
Select a bit-width to see how the number of available integer levels changes, then use the speed slider and play/pause to animate weights snapping onto the quantized grid, or rebuild to generate a new random weight distribution.
Bit-width select, speed slider, play/pause, rebuild
Google's TensorFlow Lite and Apple's Core ML both default to INT8 quantization for on-device models, shrinking many networks to roughly a quarter of their original size while running several times faster on mobile hardware.
The simulation visualizes a layer of floating-point weights being mapped onto a discrete integer grid, showing how scale and zero-point compress the continuous distribution into quantization buckets.
The simulation visualizes a layer of floating-point weights being mapped onto a discrete integer grid, showing how scale and zero-point compress the continuous distribution into quantization buckets.
Select a bit-width to see how the number of available integer levels changes, then use the speed slider and play/pause to animate weights snapping onto the quantized grid, or rebuild to generate a new random weight distribution.
Google's TensorFlow Lite and Apple's Core ML both default to INT8 quantization for on-device models, shrinking many networks to roughly a quarter of their original size while running several times faster on mobile hardware.