Results 1 to 14 of 14
05-11-12, 07:02 PM #1
Maximizing PPD for Chimp Challenge
I'm sure I'm not alone trying to find ways to maximize my output for the race. Figured I'd start a thread to ask my questions, and hopefully help anyone else who might be unsure of what/how to run.
I'm setting up 3 new rigs that I'm piecing together right now for the race. All 3 will be naked dedicated rigs: 1x Q9550, 2x Q6600.
As we all know, Linux is where it's at for PPD. I'll be overclocking all weekend, and once stable installing Ubuntu 10.10 on them all. My question is, what client will give best PPD/WUs? v6 or v7?
Feel free to post your own questions in this thread, maybe this can turn into a good resource for all of us.
05-11-12, 07:15 PM #2
6.34 console client is the best. The V7 linux client is still being "polished" - it has received less attention, than the Windows client, so it's a bit behind.
On the plus side, it has the "download the next wu at 9X%" feature built into it, so that's a good start. (Where X is any digit you select).
05-11-12, 07:20 PM #3
From what I've gathered so far, go with v6.34 under Linux for max ppd.
Also, ram-wise go for high mhz, tight timings don't help as much.
CPU only smp is the way to go too.
My two cents...
Edit: Lol at my slow typing speed...
05-11-12, 07:33 PM #4
Well then, sound like I'll be going with 6.34.
Next question. I've found my max on a Q6600 today. 3.71GHz Prime stable 24 hours. Running 411 x 9 1.58 vcore. Ram is 1066MHz. Can I save time by just using a divider to bump the ram's speed up, or should I find my max FSB? Will there be a TPF difference compared to, say, 463 x 8? In my 775 days I always ran 1:1, so I don't have much experience with dividers.
05-11-12, 07:35 PM #5
And, enter Adak and his knowledge to shed some light on that.
I think there's always a bandwidth hit when going with dividers, and you should normally go for higher ram speed than fsb -with modern ddr3 at least- but with ddr2 and prior stuff, moar fsb 1:1 was the way to go.
05-11-12, 07:42 PM #6
05-11-12, 07:51 PM #7
05-11-12, 08:34 PM #8
Fever, if you take your rig's overclock right to the max, it will fail when it's time to race. Prime95 is good, but it's not as stressful as folding.
My suggestion is to forget the dividers, and get your max stable overclock - then back it down one notch. In a 10 day race, you won't have time to recover from a 2% faster rig, that fails every third or fifth wu, losing thousands of points.
It's the old "Titanic" lesson. The "icebergs" of crashed wu's are "in the water". If you want to get to New York in a hurry, slow down a bit.
05-11-12, 10:04 PM #9
GHZ is the final key, not FSB speed.
05-11-12, 11:30 PM #10
Right now I've started testing the Q6700 at 4:5, which brings me just shy the RAM's rated speed of 1066.
Time to start overclocking the Q9550 Rex combo while the other rig is stress testing. Should be interesting, as this will be the first time I use DDR3 on this old platform. It's definitely getting hot in here!!
05-11-12, 11:57 PM #11
Also, concerning stability, I'd suggest everyone use Stanford's official F@H stress testing program before the race to make sure everything hums along nicely during the race.
05-12-12, 02:24 AM #12
05-12-12, 04:22 AM #13
Always big packet size. I don't use the advmethods or beta flags because when a work unit errors out for any reason, I would usually miss several hours of folding, before I could fix the problem.
Which is better right now? That's a question best answered by ChasR or Macaholic, who are mods for Stanford's folding forum, and more up on the latest news than the rest of us.
05-12-12, 08:24 AM #14
On the C2Qs, folding with the videocards, if you have them, is the way to maxppd. A GTX260 will make about 8000 ppd while only knocking about 1500 ppd off the smp instance. My recommendation on a 2600K would be different. It's a real PITA if you're going to be running native Linux to get the video cards going. If you have enough Windows licenses, Installing Linux in a VM and running the Video cards in Windows is the best option. By overclocking the video cards in windows (can't be easily done in Linux) you can more than make up for the loss of ppd from the VM overhead,
THe C2Qs aren't going to perform much differently regardless of dividers. My Q6600s @ 3.2 make virtually the same ppd @ 400 x 8 as 356 x 9. I don't run any of the Q6600s past 3.4 because of the voltage required to reach that. I don't think a vcore of 1.58 is sustainable, but you could prove me wrong.
I've tested ram extensively on C2Qs and found the difference between 800 4-4-4-12 and 1067 5-5-5-15 to be less than 2%, the later being faster.
I'm not currently folding enough instances to know what the "best" flags are.
OCCT is currently the best FAH stability tester.