can CUDA accelerate sb?

[jcoveiro]

can CUDA accelerate sb?

[vjs]

The easy answer is no.

I think this is due to the single point nature of the cards and programing.

You might want to post this message on the mersenneforum software section you will get more answers there or maybe already find the discussion.

[jasong]

I don't know much about CUDA, but this single-, double-precision stuff is driving me up the wall. There is no possible way that people could know that single-precision is so bad it's not worth considering unless they've done at least a tiny bit of research.

The argument goes something like,"The way the algorithm works, converting it to single-precision would make it impossibly slow." The only problem with this argument is that the algorithm itself is a shortcut made SPECIFICALLY for double-precision. There is no reason on God's green Earth that we should need to do things exactly the same on all hardware. Converting the algorithm would make it impossibly slow, absolutely. So we need to DITCH THE ALGORITHM. The algorithm isn't the mathematical theory, it's a specialized program that uses tricks to do the math faster. But if the tricks quit working, then you simply need to look for new tricks, not just totally give up.

[alpha]

That's OK in theory, but to "ditch the algorithm" is expensive. Who's going to write the new one? Where does the time and effort come from? Who's paying for it?

Writing a new, optimized algorithm for this stuff isn't something you can do in a couple of hours. If it was easy, someone would've done it by now.

[vjs]

jasong is right.

We would need a new program. On thing that I always found interesting would be the combine the GPU with the CPU.

Perhaps something could be done with P-1 for example, run B1 on the GPU pass the residual to the CPU for stage 2. Have some sort of cross talk etc where when the GPU finishes the first stage the CPU would finish the second and pick up the new stage 1 residual.

In either case you would need to find someone who knows how to program in CUDA. Personally I don't know even close to enough to begin.

[jasong]

That's OK in theory, but to "ditch the algorithm" is expensive. Who's going to write the new one? Where does the time and effort come from? Who's paying for it?

Writing a new, optimized algorithm for this stuff isn't something you can do in a couple of hours. If it was easy, someone would've done it by now.
That's true. Someone would have already done it by now. But...

How do you know someone isn't already working on a client? We might assume that that person would want to jump online and brag about what they'd done, but not everyone thinks that way. Perhaps someone has a group of small friends beta-testing and that someone likes to talk about what they're doing to a like-minded person online.

I don't have the skills to attempt or understand what's needed, but I DO have the ability to note someone's reputation online and develop an opinion of what they're knowledgeable about.

Sorry for the cryptic message, I guess I'm a bit like a worm teasing a fish because the cat isn't quite ready yet. ;) Give it another few months, and we'll see what develops.