UD-Q3_K_XL: 80.7% (−0.6 points; +3.5% relative error increase)
并且这些发光像素可以自由控制，S26 Ultra 的防窥模式还拥有两个防窥膜永远无法达成的效果：
，这一点在PDF资料中也有详细论述
Why does this paper have a chance of succeeding?
This isn't as much as an issue to me since I pack vertex data into bytes with the format of SDL_GPU_VERTEXELEMENTFORMAT_UBYTE4. Even when I used doubles previously, they only existed when copying chunks into the GPU (Dyalog will convert them to floats if they're passed to a C function that takes floats). However, if you're doing anything that would require having data that would usually be stored as floats in memory (maybe for updating every frame?), then this might be an issue.