zeckensack said:
The real reason you want to do it is because it achieves pretty much perfect load balancing. That's not to say that I think it's easy to pull this off. But once you do, this is the benefit you're going to get. It's all about performance, not about features or the programming model.
I totally agree with you, but there is definately a connection to features, namely vertex texturing and dynamic branching.
Current games are not very vertex heavy, so IHV's can get away with fewer vertex units. Add the fact that their operations are pure arithmetic, and they don't need much latency hiding, resulting in compactness. I don't think there is that much space to save by eliminating the VS, so generally speaking, I think gains of a USA in today's games will not be very big.
It's tough to say, because I know current architechtures have a big FIFO between the projection unit and PS to minimize the impact of small/zero-pixel triangles clusters, and this can be nearly eliminated with good load balancing. On the other hand, the additional data routing and scheduling needed is big, so that would likely take up plenty of space.
Throw vertex texturing into the picture, and suddenly you vertex units must be a lot bigger if they're going to perform well. Good dynamic branching needs a good scheduler, so you probably don't need a whole lot more to throw vertex management into the mix. Now USA makes a lot more sense, since the die savings are bigger, and thus you can add more ALUs like Xenos did.
So yes, it is all about performance, but performance depends on the load, which depends on the features used, which depend on the shader model and how developers use it. So in the ends, it's not really
all about the performance.
BTW, many of the explanations above are not directed at you, since you're probably more knowledgeable than I am. They're for other members' benefit.
