I was only vaguely aware of this thanks to graphics programming, where write-combined access to GPU memory means you'll be more than twice as fast by writing en

I was only vaguely aware of this thanks to graphics programming, where write-combined access to GPU memory means you'll be more than twice as fast by writing entire cachelines without holes (because reading write-combined cachelines is effectively uncached? Horribly slow.) Documentation for 3D rendering APIs tend to have warnings about this as a result, and performance talks tend to cover it as well. "memcpy the (cachline aligned and write combined) site from orbit - it's the only way to be sure"