CPU Architecture and FAH

[7im]

Buy the fastest processor with the more cores that you can afford. Yes, quad cores will increase your PPD quite a bit, while not increasing your electricity cost that much.

[shatteredsilicon]

So, if I'm following this correctly, you are asking what CPU will be sufficient to saturate whatever number of GPUs you are planning to have. If that's the case, then I can confirm that my own experience indicates that it takes approximately 15-20% (depending on the WU) of a 2.6GHz+ Core2/Phenom class CPU to keep a G92 GPU saturated. G200s may take a bit more. So, if you are planning to run <= 5 GPUs (probably 4 if you are running G200s) in the system, a Core2 Solo would be sufficient to keep the GPUs fed (you'd have no CPU time left for an SMP client, though). This is under Linux with the wine cuda wrappers, and a kernel with high-res timer support.

[bruce]

Welcome to the foldingforum, Archangelboy.

The FAH software does it's best to take advantage of whatever hardware you have. Clearly a Quad can process twice as many instructions in a given time compared to a Duo so (assuming other factors don't change) you'd expect nearly twice the PPD. The CPU will cost more and will draw more power, but the RAM, Video, HD, etc. will not change so you'll get a more efficient system.

Comparing the Core2 architecture to older intel architecture, they did improve the efficiency of Floating Point (and SSE) math, so yes, it's more efficient than older machines. Actual FAH speed improvements comparing i7 with Core2 are small, although HyperThreading can help some non-FAH code quite a bit.

GPUs are able to process a lot more instructions in parallel than today's CPUs, so they get higher throughput.

As far as which architecture the Pande Group would prefer, they are thankful for anything you're willing to contribute to the project (Duh). If you're willing to invest more money or more electricity in newer/faster technology, that's even better, but there's no simple answer to your question. They'd also be happier if you fold on a dozen computers instead of just one or two.

Let me turn the question around another way. If you have a computer that could be folding but isn't, why isn't it?

[Archangelboy]

Thanks, Bruce, and well put!

Just picked up a(nother) used i7, plan on putting it and a 295 online as soon as I can afford RAM and a PSU in addition to what I've got running now.

Any word on how i9 Intels will perform? similarly to i7 (with scale-up for additional cores) or will there be an improvement/core (or per client) with that architecture?

[toTOW]

See this news on FAH-Addict : first review of Core i9 (yes I know, it's in French ... or Chinise for the review )

[Archangelboy]

C'est d'accord, toTow, j'en parle un peu. Merci

[divery4eyes]

sad to say, but I do not foresee AMD being with us much longer

[Archangelboy]

Lol nah, if AMD can't compete, they'll just whine until somebody makes Intel pay to keep their operation afloat. Thank you, EU >.<

[shatteredsilicon]

sad to say, but I do not foresee AMD being with us much longer

We can but hope. Then maybe we can finally knock x86 on the head once and for all.

[YashBudini]

My question would be in regards to quad cores and SMP under Windows. We've seen that the 4th core is less busy, and some people have dealt with that in various ways, like running 2 iterations of SMP on a quad core (I'm not advocating that nor do I want to start a discussion about it.) My question? Given Windows behavior and disregarding the L3 cache differences for the moment, would an AMD X3 710 (3 cores @ 2.6GHZ) fold significantly slower (25% or more) than an X4 810 (4 cores @ same speed)? The idea being to keep all cores as busy as possible, presumably when the threads outnumber the cores?

I'd expect the 710 to be slower than the 810, but it would be interesting to run a test and see how much.

Last question - does this Windows inefficiency still exist in Win 7?

[John Naylor]

The windows inefficiency is related to the design of the core rather than the OS. However if the OS could affect it, it is unlikely it would be fixed by Win7 as Win7's kernel is a development of Vista's kernel.

[MtM]

John if the os was not a problem a port from linux a2 to windows a2 wouldn't take this long. There are inherant issues with the scheduler which prevented a quick port, but I'm not aware of the specific details. Maybe 7im or Bruce can elaborate ( or correct me ) on the issue?

@yashbudini it's not only relevant how high the utilisation of each core is, it's also very relevant to how balanced the execution of each thread is in regards to the others. This is also the main argument toTow focuses on with his view on the current linux 2.10 core issues. If one thread has to wait for the others it slows down the computing process, so while in theory a tri core could offer the same performance for each thread as a quad core ( assuming about 75% usage on the quad which isn't exactly true but let's use it for arguments sake ) the synchronisation between the threads would negatively impact the final performance.

That's in theory, in practise I'm fairly certain people have tested tri vs quad core but I'm not one of them and can't recall from memory what the outcome has been ( or for that matter at what forum the thread was which I read about it, I am almost certain it was not here but probably on one of the russian/ukraine forums. I think Anglik666 was the one who lead me to that forum so it might be the polish team forum as well ).

I'll send him a pm with link to this thread, I think he can say more about the testing and results

[John Naylor]

I meant that a1 had/has efficiency problems on all 3 platforms, so win7 is unlikely to fix these problems not referring to a2 lol

EDIT: Also afaik the issues with the a2 core were down to Gromacs 4.0 compiling on Windows but not producing accurate results... since we know SMP2 development is now underway I doubt they would bother to try to port a2 now, even though the existence of the released core a4 shows that gromacs 4 now produces correct results on windows.

[Anglik666]

Well, this days I have had Radeon HD and Phenom CPU, GPU's taken 100% of one core so 3 left for SMP.
Using Mpich (for win with core a1) I've got almost twice worst PPD for 3 cores then for 4, so I used this one: viewtopic.php?f=14&t=5564.
I was sorting processes by mem, cpu usage, rotating them, made the magic 12 pairs of cpus for 4 processes etc.
The best effect I've got sorting them by mem (randomize like) and for 12 pairs. It still was more then 30%-40% worst then 4 cores but not 2 times.
After that I found the way to run Deino in 64bits M$ systems and it gave me results only 30% worst ish without any 3rd party program at all.
So in my opinion Deino much better scales itself.

But, remember, I've never checked that with stopper, always more or less.

[bruce]

John if the os was not a problem a port from linux a2 to windows a2 wouldn't take this long. There are inherant issues with the scheduler which prevented a quick port, but I'm not aware of the specific details. Maybe 7im or Bruce can elaborate ( or correct me ) on the issue?

The only think I've heard is that A2 does not work on Windows. I don't know the details of why, but since it's apparently "impossible" to port it, you're wrong about how long it "should take."

I agree with John.

Also afaik the issues with the a2 core were down to Gromacs 4.0 compiling on Windows but not producing accurate results... since we know SMP2 development is now underway I doubt they would bother to try to port a2 now, even though the existence of the released core a4 shows that gromacs 4 now produces correct results on windows.

I expect that A1 (and maybe A2, too) will be replaced with whatever code they're talking about when they mention "SMP2" but there's no way to predict when that might be.