leechan25 said:
I'm going to address this multi-core GPU thing head on because I think a lot of people are unclear on the whole difference between GPU and CPU architectures, and leechan this is not specific to you.
CPU's exist in a world where right now, parralelization of code (ie threading) is not all that common, and in the instances it *does* exist, there are diminishing returns as you increase the number of cores.
Ok - so why go multi-core then, right? Well, because we've kind of hit a wall of sorts in terms of processor speed. But we all know this and that whole story, so I'm going to skip over it.
Anyway, your typical multi-core CPU is two identical cores for the purpose of increasing (doubling) compute power in a reasonable fashion. Obviously it costs you roughly two times the die area of a single core, and there are different ways to implement it; a la Athlon vs Smithfield. Also cache can be divied up in different ways, but we're not going there either.
So let's go to GPU's. Instead of a situation where threading is difficult, on the contrary you have a situation where it's inherently crazy parralel in nature. Such that, cores have been 'multi' core for years. What does a dual core NV43 look like? Well it looks a lot like an NV40 to me (sure sure ROP and clockspeed differences, bare with me people). Want to double the performance of a GPU? Why add another core when doubling the pixel pipes, vertex shaders, ROPs, etc.. results in the exact same thing? Only better in many cases. In fact to double those components you're actually adding many cores to begin with. Now with GPU's how much benefit you derive from those extra 'cores' will depend on memory bandwidth and a whole bunch of other junk we won't get into here, but the simple matter is when people generally hear talk of multi-core GPU's, seriously the instinctive reaction is either 'why?' or 'we already have that.'
The only reasons one would actually do a true dual-core GPU would be if the targeted performance goal requied such a massive transistor investment, that yields on the resultant die size would simply be infeasible. Then what you would be left with would instead be two smaller GPU's in a sort of on-package super SLI. I don't know what to say to that, except that I know RSX in PS3 does *not* have that sort of transistor budget allocated to it.
If one wants to discuss the merits and benefits and reasons to go dual-core on the GPU, that discussion is happening right now in the G71 speculation thread:
(which has turned into G80 speculation)
http://www.beyond3d.com/forum/showthread.php?t=26270&page=38
Last edited by a moderator:
Cause it doesn't risk to slow down the GPU instead? The "second video processing" would have to be processed somewhere, moving it from CPU to GPU would free the CPU but might slow down the GPU instead, so the game on you first monitor would still risk to be slowed down.
