Each column of glowing bars is a small slice of a neural network's weight matrix. Bar height and color encode a weight's value. Every value starts as a 32-bit floating-point number — but storing and multiplying billions of those is slow and memory-hungry. Quantization rounds each weight onto a small grid of evenly-spaced integer levels (as few as 21 for 1-bit, up to 28 for 8-bit), then remembers only a scale factor to map the integers back to approximate real numbers.
INT8 quantization is the industry default for deploying large models on phones and edge devices: it typically shrinks a model to roughly a quarter of its FP32 size and can speed up inference several-fold on hardware with dedicated integer math units, often with well under 1% accuracy loss when done carefully.
A grid of glowing bars represents a slice of a neural network's weight matrix. Each bar's true 32-bit value hovers as a translucent wafer; lower the bit width and watch the bars snap onto a coarser grid of integer levels while the gap to the true value — the quantization error — grows.
Affine quantization maps a continuous range of float weights onto a small number of evenly spaced integers using a scale (and optional zero-point). Fewer bits mean fewer levels, more rounding, and higher error — but far less memory and faster integer math.
Drag the bit-width slider from 8 down to 1 and watch the bars quantize. Switch weight distributions to see how outliers wreck low-bit accuracy, try per-channel scaling to fix it, and toggle the FP32 baseline to compare against full precision.
INT8 quantization is the default for deploying large models on phones and edge devices — it shrinks a model to roughly a quarter of its FP32 size with often under 1% accuracy loss.