Seed some information said that the architecture of GT300 will be as MPMD( multi-program multi-data).
It`s true?
I can guarantee that people who know the answer to this question will not be allowed to comment. :) (I’m not one of them, so don’t look at me…)
I read about it a while ago. As seibert said, I guess no one could provide exact answer here. But if it’s true, that would bring GT300 a huge benefit over other up-coming GPUs like RV870 and Intel Larrabee. Especially for Larrabee, which is designed with CPU mind-set and uses vector-based instruction, may have lower overhead to dynamic branching than current GPU architectures. In addition, this is my personal guess, the GT300 might (or should?) have another plan to improve memory coalescing (just like what they changed in G80/G92 => GT200) because it’s critical to many applications and Larrabee claims to use large (or smart?) cache and ring bus to minimize the memory access latency (though nobody except Intel knows how good it is until they launch it next year). But clearly the memory access pattern optimization is a headache for most CUDA developers and Intel seems well-prepared for it when throwing out Larrabee, so…anyway let’s see! :rolleyes: