BUT...I saw this the other day on PrimeGrid:
"For the most part, all of the software that we use is highly optimized and hand-tuned for the particular hardware it's designed to run on. The best example of this is George Woltman's gwnum libraries, which is inside LLR and a lot of other software. It's hand-written assembly language code designed specifically for the various CPU improvements that have been created over the years.

It squeezes every last drop of performance out of whatever CPU you have.

Opening up the CUDA architecture to more platforms makes it easy to write portable code. It won't necessarily make it easy to write fast code for all platforms. You can have portable code, or you can have fast code, but fast portable code is almost an oxymoron.

Take a look at our GPU sieving software, for example.

On Nvidia GPUs, the software is written in CUDA, which is sepcifically designed for the Nvidia CPUs. It's blazingly fast.

On ATI/AMD, it's written in OpenCL, which is portable and is not specifically designed for ATI's GPUs. It's pitifully slow by comparison to the Nvidia version of the software. The hardware itself is not inferior; it's the fact that a portable, cross platform software architecture is being used that makes it so slow."

Unfortunately there are tradeoffs for all 'generic' software.

Why HCC already runs on CPUs?

Highly unlikely. No portable code is ever as fast as that written for the specific platform it is to run on.

just waiting and wondering what will come up next till then i will wait

If an OpenCL app will be used for CPU's, the app is either as fast, faster or there are scientific research reasons (similar apps make for more comparable and reliable results). I mentioned this in another GPU thread.