LL with OpenCL

[kracker]

Hate to break it to you, but LL tests are very unscalable, and bruteforce style will not work. And I can just imagine the latency of lan/vs PCI-E... LL tests aren't something distributable (etc half of work on one gpu and half on another) each iter depends on the previous one.

[Robish]

Originally Posted by Karl M Johnson

Multi-gpu support should be added first then.
It's not an easy undertaking, even if a single thread controls all the gpus.
It makes more sense to add multi-gpu support to CUDALucas/clLucas, since it would ideally cut the LL test time by the number of GPUs in the system (if they are the same and versus a single gpu).
Once that is done, merging several rigs and assigning a single exponent to them could be considered.

That's my thinking exactly, I agree it probably would be difficult (way beyond me) but Atom from OCLHashcat has done something similar so there's a starting point for an approach. and the benefits could be enormous.

[Robish]

Originally Posted by kracker

Hate to break it to you, but LL tests are very unscalable, and bruteforce style will not work. And I can just imagine the latency of lan/vs PCI-E... LL tests aren't something distributable (etc half of work on one gpu and half on another) each iter depends on the previous one.

Point taken, I just thought the discussion was worth exploring

[VictordeHolland]

It's not all about computing power, memory bandwidth also plays a big role. PCI-E is quite fast, but not as fast as the GPU memory. Connecting >3 GPUs will also require very expensive motherboards and infinityband (or similar). Running one task on each GPU is probably going to be much more efficiënt and less expensive.

[Robish]

Originally Posted by Robish

Point taken, I just thought the discussion was worth exploring

Its just when you read articles like this you wonder if things can advance - dramatically.

http://arstechnica.com/security/2012...rd-in-6-hours/

"350 billion-guess-per-second " I mean...WOW that's powerful. But I take on board Krackers' comments. I don't know enough about LL to even argue ;-)

[kracker]

Originally Posted by Robish

Its just when you read articles like this you wonder if things can advance - dramatically.

http://arstechnica.com/security/2012...rd-in-6-hours/

"350 billion-guess-per-second " I mean...WOW that's powerful. But I take on board Krackers' comments. I don't know enough about LL to even argue ;-)

Well, it's not like LL tests take a humongous time to do these days, anyways.

(Wish I had a GPU setup like that)

[Robish]

Originally Posted by VictordeHolland

It's not all about computing power, memory bandwidth also plays a big role. PCI-E is quite fast, but not as fast as the GPU memory. Connecting >3 GPUs will also require very expensive motherboards and infinityband (or similar). Running one task on each GPU is probably going to be much more efficiënt and less expensive.

You're right. Infiniband is what Gosney uses on his rig. 10 Gb/sec. However he says he bought all cards connectors and switch for $800 on ebay (used) - less than the price of a single high end GPU.

okay I'll shut up about it now, just thought it worth mentioning.. :-)

[Robish]

Originally Posted by kracker

Well, it's not like LL tests take a humongous time to do these days, anyways.

(Wish I had a GPU setup like that)

:-) i'm only a newbie here Kracker, did it take longer a few years ago? I suppose it had to of. Older chips n such.

(trying to build a smaller scale one) Learning Linux is tough (no windows support) but I'm getting there :-)

[Robish]

Originally Posted by Robish

:-) i'm only a newbie here Kracker, did it take longer a few years ago? I suppose it had to of. Older chips n such.

(trying to build a smaller scale one) Learning Linux is tough (no windows support) but I'm getting there :-)

I'm sorry did I get distracted....Breaking Bad ROCKS!!!! :-) hic ahem sorry eh er ahem cough, bit drunk right now hmmm ah ill be ok tomorrow and do more LLs :-)

Ps prime searching is fun is it not? Whiskey however......more. ;-)

[kracker]

@msft: by the way, is there a "worktodo.txt" in clLucas right now?

[Karl M Johnson]

Originally Posted by kracker

Hate to break it to you, but LL tests are very unscalable... LL tests aren't something distributable (etc half of work on one gpu and half on another) each iter depends on the previous one.

Indeed.
Mlucas' parallel code doesn't scale well beyond 4 cores, but there's still an increase in speed versus 4 cores.
There will be no ideal 300% increase in speed for, say, 4 GTX Titans or 4 7970s, but the status of a single potential Mersenne number will surely be known sooner (versus a single GPU).
Besides, the availability of several gaming GPUs inside a single machine is higher than of several server class CPUs.
The idea has a chance to live.