Draco, Meshopt, KTX2, and the texture memory you forgot to count

Draco compresses mesh data hard, often to a tenth, at the cost of a WASM decoder and a decode pause on the main thread. Meshopt compresses less but decodes an order of magnitude faster and streams better. For a hero model behind a loading screen, Draco. For many small assets loaded during interaction, Meshopt. Choosing by which one has the better compression ratio is the wrong criterion.

Textures are the real budget. A 2048 by 2048 PNG might transfer as 1.5MB, but on the GPU it is decompressed to raw pixels: 2048 times 2048 times 4 bytes is 16MB, plus a third again for mipmaps, so about 22MB. Four of those and you have spent 88MB on a device that may only give you a few hundred, and the first sign of trouble is not a warning but a tab that dies on an older iPhone.

KTX2 with Basis Universal is the fix, and it is the least-known important thing in this phase. It transcodes at load into a format the GPU keeps compressed in memory, typically a quarter of the size, permanently, not just in transit. It looks slightly worse under close inspection and it is almost always the correct trade.