Adi Oanca <adioanca@xxxxxxxxxxxxxx> wrote: > > I'm actually not sure if that assumption is true - it depends on > > how > > you draw, and that can be optimized for this case. Video memory > > should > > be mapped in "write-combined" mode which makes sure that the writes > > leave the cache as 32 byte bursts. > It's not an assumption. Telling the graphics card to take a piece > of data from > main memory is faster (uses AGP) than writing 32 bytes times and > times over. I'm pretty sure you're wrong here. A bus is a bus - ie. the graphics card cannot access the memory bus faster than the CPU can; the only advantage the GPU has over the CPU is that it can ask for more than one piece of memory at once, it doesn't have to wait for the data anymore (with AGP). Why should the bus be slower for one component when it does exactly the same work? But since the CPU's caches work independently from the bus frequency, this shouldn't matter that much - it all depends on how you fill the data in memory, though, if you leave the cache frame, or even those 32 byte array, you'll definitely lose performance over fetching the data from the GPU. The bus memory bandwidth is precious for graphics, and you can easily fill it up by having a wrong design in the app_server during video playback or something like this, and this will hurt overall performance. Bye, Axel.