MoTheG (better@not.tell) on 11/8/09 wrote:
---------------------------
>I still am surprised that the Atom scored so well.

I was as well when I was putting this together. Initially I just had high performance CPUs, and I thought for fun I'd throw in GPUs and Atom and see where they stand. I was more than a little surprised at the results.

>ofcause the benchmark was parallel and explicit at that, >thus there was no gain
>from OoO-processing but a factor of >3 ?

Well there wasn't a benchmark, but you are absolutely correct the metric didn't incorporate the advantages of OOO. Or caches for that matter.

>Doesn't the article make a good argument for keeping CPUs >and GPUs different?
>A CPU can not be as good at large parallel tasks because >it carrys its sequential strength as a burden.

Actually that strength really is a strength. Many workloads are very difficult to parallelize or vectorize with a compiler. Think GCC or most latency sensitive applications.

There's no reason why a CPU cannot have a bunch of massively parallel execution resources like a GPU and simply use the most appropriate hardware for the task.

From a physical perspective, the distinction is really about power consumption and memory bandwidth vs. latency.

A separate GPU makes a lot of sense where you need lots of power and memory bandwidth AND the CPU guys don't want to provide it. As you can see from Intel and AMD, they are now interested in figuring out how to work with the folks who do have those needs.

>with the GPU getting easyer to use (OpenCL, DirectX11, >CUDA, Brook+) and better
>connected (PEG), CPUs should concentrate on those tasks >that are to small to be off-loaded.

GPUs are getting easier to use means they are improving relatively speaking to prior generations. It says nothing about suitability for the general programming populace.

Explicitly parallel stuff is easy to exploit, and if you're going to rewrite software, then you need to compare optimized CPU code vs. GPU code. That's something NV has a big problem understanding - most of their magical 100X speed ups are probably comparing poorly optimized CPU code with optimized GPU code.

David