Aaron Spink (aaronspink@notearthlink.net) on 8/2/10 wrote:
---------------------------
>hobold (hobold@vectorizer.org) on 7/30/10 wrote:
>---------------------------
>>Given that GPUs have only just begun to delve into the same kind of microarchitectural
>>sophistication that CPUs have harvested for a few decades (Nvidia's GF104 is the
>>first superscalar GPU), I would like to think that GPUs do have several promising
>>avenues to follow. CPUs in turn are left with few options beyond more cores and wider SIMD.
>>
>You do realize that what you just said was: GPUs are limited now to doing the hard
>stuff. CPUs still have all the low hanging fruit to exploit.
>
Our views are not as contradictory as you seem to imply. For the CPU guys, more cores/threads and wider SIMD are problematic because they are visible to software. But hardware-wise, they are easy to do.

For the GPU guys, many cores/threads and wider SIMD are easy, because the programming model is transparent to them. Superscalar dynamic OoOE will be hard to implement in the GPUs, albeit some of the CPU know-how can be leveraged. But the programming model will again be totally transparent to those hardware techniques.

So I was looking at this from the bottleneck that is software production, while you are looking at it from the bottleneck that is hardware production.

>Its rather easy to add more cores and increase SIMD widths and CPU have the software
>ecosystems in place to quickly exploit the increased cores and increased SIMD widths.
>
CPUs are only slowly crawling towards that kind of software ecosystem. Unless you use OpenCL for CPUs. :-) But they will get there eventually. They have no other way to go.