Another problem is immediate mode vertex data is tied to the GPU front end vertex / data fetch rate, while vertex buffer data is read through the texture fetch path (much, much higher bandwidth). So the max rate for each type is fundamentally bottlenecked by the architecture, even if you could transfer it in at the same rate. The fixed function GPU front end is not as massively parallel as the general purpose shader engines / texture fetch portion. To reach the same levels of performance on immediate mode, the architecture would have to change significantly, and it would most likely come with a large area hit, which directly translates into chip yield / cost.

I think the part of the deal where AMD can go fabless and not have to manufacture x% of their x86 processors in their own fabs is worth just as much as the cash. That lets them free themselves of Global Foundries and become an innovative product company without having to worry about the fab boat anchor. Sure, they won't be able to write their own design rules and tweak the last 5% performance or area out of the process, but they also won't be in mortal danger when the fab isn't running at capacity. Intel can afford the fabs at this point, and AMD has restructured itself to not be able to afford them. Hopefully they can make up for it in innovation, time to market, "agility", ...

I think the chatty paradigm of GPU usage will be more fine-grained "stream computing." When the latency between CPU and GPU is lower, and you share the same cache, the penalty for setting up and launching stream computing tasks on the GPU becomes lower, enabling more things to be accelerated this way.

The old way, you only really got benefits from stream computing if you were able to set up a massive job for the GPU, set it on its task, wait for completion, and then get the results. Now, maybe new classes of apps become more feasible.