Block compression
Lossy, fixed-ratio texture compression in which 4×4 blocks of texels are packed into fixed-size blocks that the graphics card decodes directly at sample time.
An ordinary image has to be fully unpacked before it is used. Block compression is instead built so the GPU's texture sampler can decode a single 4×4 block in place — the texture stays compressed in video memory, not merely on disk. Each block is coded into 8 or 16 bytes, so the ratio is fixed and memory use is predictable. The price is a small, permanent loss of quality that is hard to notice in motion.
The technique arrived as S3TC/DXTn (S3 Graphics, 1998) and was standardised as BC1–BC7 in DirectX 10/11. BC1 gives 6:1 on opaque textures; BC7 spends more bits at the same size for better quality. The mobile equivalents are called ASTC and ETC. Pair it with mipmaps to keep both bandwidth and the cache in check.