DDR Bandwidth and QMD. need help choosing new memory

[davekusa]

Looks like the QMD respond well to Memory bandwidth. I may have a line on some dirt cheap Pentium 3.2 Processors and Might be looking to double my farm.

I have Corsair XMS 3200 2-2-2-5 TCCD.
Sandra shows about 5800 bandwidth 1:1 at 240 2.5-3-3-5
Is there a faster memory out that I can try.
I've been scouring the web and havn't found information thats cut and dry.

Any help would be appreciated

[consumer9000]

Get a system with DDR2 support. My fastest QMD folding setup has 2048MB of Samsung PC4300 DDR2 memory - the enhanced prefetch etc and the 925X chipset seem to be helpful.

[TollhouseFrank]

I think most of the WU's love better Bandwidth. Just by increasing from 166 to 200 in BIOS (a SMALL overclock).... all my wu's fold about 10-15% faster!

Bandwidth is king!

[davekusa]

I got a deal on a 478 chip. I think only 775 supports DDR2.

[KrisMCool]

Your TCCD should do well. Yes, QMD WUs respond well to better memory bandwidth and timings. If you look at the rigs in my sig, you might guess that the Max3 rig at FSB 250 would be faster than the AI7 rig at 234, roughly equal CPU clocks. But the AI7 has been consistently faster, at least partly due to 2-2-2 timings. I tested the AI7 with other memory types and timings, and folding times got noticeably slower with looser timings. The only other memory I can think of ATM that would give you better timings at 240 is Micron based (2.5-2-2); other S478 P4 users have had good results with these at those speeds. I wouldn't suggest BH/CH UTT memory as it's a real crap shoot as to whether it will work with your P4 rig; my UTT based memory hates my P4 rigs.

[ChasR]

If you have enough bandwidth to fold one QMD more bandwidth won't make it fold faster. More bandwidth will make two fold faster but there isn't currently an Intel rig out there with enough to fold two without running into a bandwidth bottleneck. As a result, you won't see much of an improvement from hyperthreading. Latency is the key. OCZ VX which is Winbond based memory is the only memory I know of thats faster than the TCCD memory. I've tried the Micron based memory and found it to be inconsistent, VX has to have a lot of volts (3.3 volts) but it will run at 2-2-2-5 past DDR 500. Frank, the effect you're experiencing in all WUs is not due to the increase in memory bandwith but due to the lower effective latency that comes with the higher FSB.

[TollhouseFrank]

so that's what It is? Mind explaining it to me? I don't quite understand the point you are getting at on the lower latency ChasR... my memory settings didn't change.... how could the latency be lower? Or am I misunderstanding what you are saying?

[ChasR]

CAS 2 means that the data is valid on the rising edge of the second clock after READ is issued. If you speed up the clock (raise the FSB) the amount of time between the rising edges of the sine wave is shortened thus the effective latency is decreased.

Memory latency is expressed in clock cycles. Effective latency in nanoseconds. 2 x (1/333,000,000)sec is a longer period of time than 2 x (1/400,000,000)sec

[TollhouseFrank]

ah... OK....

Thanks ChasR

[Joe Camel]

huh, guess that explains why:
my 3500 Venice @ 2.7GHz [270 x 10 (TCCD @ 2.5-4-3-7)]
was only a little faster than:
my 3200 Winnie @ 2.3GHz [230 x 10 (OCZ VX @ 2-2-2-7)]
on the same 100 frame Gromacs (364 point) WU... around 35 sec faster/frame (about 1 hour on a (+/-) 24 hour WU)

note: these are NOT exact #'s, they are from memory (mine) which is...shall we say...not the best :rolleyes:
next time they are on the SAME WU, ill take notes ;)

[Silver]

On the 2.8c I ran the ram at 2,2,2,5 using a divider for 3.5Ghz. Added a OCZ booster a couple of weeks ago, lowered the oc to 3457Mhz and ran the ram at 2,2,2,6 at 1:1. I have been running this setup for the last two weeks after trying 250 fsb 1:1 with 2,2,2,6 (would fail half way through a QMD). The present settings at 247fsb with the tight timings and 1:1 has increased my PPD on this unit even though the overall oc is lower. Was getting 466PPD on the system and am now getting 544 PPD on the system as reported for the last two weeks. Not too bad for a 2.8c. I do believe that 250 1:1 would be possible at 3.5v on the ram but need better active cooling to attempt it. During the half completed QMDs, PPD was reported at 566 using these settings. These results have caused me to rethink the mods I was contemplating for the xeons as well as the dual core that I would like to purchase. Bandwidth on the xeons is not going to even come close and I have to wonder about the new dual core Intels as well. I am presently leaning towards the X2s as one would think QMDs can not be kept from them forever. One further note, the 2.8c (and 2800+) may both be limited by the 380 w Antecs powering them (though I am very happy with the results from these psu's). All voltages are stable though at 3.5Ghz this was not the case on the tight timings at 1:1.

Ram is the Corsair XMS 3200 (reported by cpu-z as 3200 but part number cmx256a-3500c2). Voltage is 3.4v.

[Joe Camel]

...Added a OCZ booster a couple of weeks ago... I do believe that 250 1:1 would be possible at 3.5v on the ram but need better active cooling to attempt it. ...
HERE (http://www.bleedinedge.com/forum/showthread.php?t=10902) is a GREAT setup for the Booster's HS

0 (zero) voltage fluctuation (booster readout) after mod :attn:

GOOD LUCK!!

[stratcatprowlin]

Hey btw,is 512mb's enough for a dedicated QMD folder that does absolutely nothing else?

[Silver]

Hey btw,is 512mb's enough for a dedicated QMD folder that does absolutely nothing else?

Works for me and in fact can hyperthread it with one qmd and anything else but another qmd. Problem is total points falls. So, running two 256 sticks and dedicated QMDs on that unit. Can surf while it is folding though it is not the box I use to surf.

Chas, what's your take on the new dual cores vs bandwidth? Think one would still have the same problem as the xeons (though to a lesser degree but none the less).

754 boxes are all running single 256 sticks and the xeon setup is running dual 512 values with one qmd and one bigpacket (non Qmd) for pointage between 460 and 630 per day depending on the second unit. 754's are pretty consistantly (now) running 330 ppd.

[davekusa]

this is some great information. So it seems that the actual memory doesn't matter. It's all about how high a FSB you can push it with tight timing.

[Silver]

HERE (http://www.bleedinedge.com/forum/showthread.php?t=10902) is a GREAT setup for the Booster's HS

0 (zero) voltage fluctuation (booster readout) after mod :attn:

GOOD LUCK!!

Looks good. Would seem to be a very good mod especially with the xp-90. Move the booster over so it is under the xp-90 zalman fan, along with the heatsink mod, take the voltage up to 3.5v and attempt once again to pull 250 2,2,2,6 1:1.

It is kind of nice that the wife has limited me to the four systems as it is now all about tweaking. Hoping to pull 1700 PPD after having tweaked all the systems again. Just need to get a weeks worth in again. I honestly believe that 1800 PPD can be pulled from the four boxes (754-3000+, 754-2800+,2.8c and dual 1.6 xeons) with just a little more tweaking, a fan here and a heatsink there while still keeping things (and electric bill) out of the extreme cooling arena.

[ChasR]

Chas, what's your take on the new dual cores vs bandwidth? Think one would still have the same problem as the xeons (though to a lesser degree but none the less).

It will behave almost exactly like the 875 chipset duallies (ASUS NCCH-DL) which have a bandwidth of around 4300MB/sec. The dual core will be a good bit better than the 7xxx chipset duallies. At present there are only two families of WUs that demand lots of bandwidth, p147x (of which few are available) and QMDs. If it weren't for QMDs the dual cores would seem like great folders and for the money they are. Folding one QMD and one big Gromac WU should produce somewhere around 675 ppd on the dual core @ 3.2 GHz. Folding two QMDs it'll be under 600 ppd.

[Silver]

Pretty much what I was thinking. I of course will wait for results but it is not looking real promising. Will of course wait for x2's to drop some but given that in folding one is not updating on a regular basis, the x2's seem to be the ones. If one looks to buy and hang on for at least a year then they will most likely get the QMD thing ironed out and the x2's should benefit greatly. At least that is my thinking as of now.

[ChasR]

Long term the x2 is the better choice but it's going to be expensive at least initially.

[Silver]

Building from the ground up might not be too bad. I did a price check and there is about $100 difference in mobo's (intel is more) then there is the 939 agp of which it is my understanding they just sent out samples. If that is true and it will run the x2 then we are really looking at roughly the same cost for a mobo, vid (reused on the agp), cpu and ram. As a matter of fact the psu may need upgrading on the Intel adding to its cost. It is interesting to note that the cost of the AMD (though more) is really giving AMD a windfall on the mobo as in the cost of the Intel dual core one must purchase the mobo and this just ups the cost to the cost of the x2 (or vis-a-versa).

[stratcatprowlin]

Look at the difference the ram amount and timings made here.Unbeleivable!

QMD P4 3.0 Prescott with 1 gig of dual channel pc3200 2.5-3-3-8

[02:18:40] Completed 1473 out of 2103 steps (70)
[02:18:40] Writing local files
[02:33:43] Timered checkpoint triggered.
[02:36:16] Completed 1494 out of 2103 steps (71)
[02:36:16] Writing local files
[02:51:36] Timered checkpoint triggered.
[02:53:09] Completed 1515 out of 2103 steps (72)
[02:53:09] Writing local files
[03:08:51] Timered checkpoint triggered.
[03:10:28] Completed 1536 out of 2103 steps (73)
[03:10:28] Writing local files
[03:26:06] Timered checkpoint triggered.
[03:27:42] Completed 1557 out of 2103 steps (74)
[03:27:42] Writing local files
[03:42:53] Timered checkpoint triggered.
[03:44:31] Completed 1578 out of 2103 steps (75)
[03:44:31] Writing local files
[03:59:39] Timered checkpoint triggered.
[04:01:16] Completed 1599 out of 2103 steps (76)
[04:01:16] Writing local files
[04:16:26] Timered checkpoint triggered.
[04:18:04] Completed 1620 out of 2103 steps (77)
[04:18:04] Writing local files
[04:33:16] Timered checkpoint triggered.
[04:34:53] Completed 1641 out of 2103 steps (78)
[04:34:53] Writing local files
[04:50:05] Timered checkpoint triggered.
[04:51:43] Completed 1662 out of 2103 steps (79)
[04:51:43] Writing local files
[05:06:52] Timered checkpoint triggered.
[05:08:30] Completed 1683 out of 2103 steps (80)

QMD P4 3.0 Prescott with one 512 stick single channel pc3200 3-4-4-8
[09:00:42] Completed 745 out of 2067 steps (36)
[09:00:42] Writing local files
[09:19:56] Completed 765 out of 2067 steps (37)
[09:19:56] Writing local files
[09:40:07] Completed 786 out of 2067 steps (38)
[09:40:07] Writing local files
[10:00:18] Completed 807 out of 2067 steps (39)
[10:00:18] Writing local files
[10:19:31] Completed 827 out of 2067 steps (40)
[10:19:31] Writing local files
[10:39:42] Completed 848 out of 2067 steps (41)
[10:39:42] Writing local files
[10:59:53] Completed 869 out of 2067 steps (42)
[10:59:53] Writing local files
[11:19:06] Completed 889 out of 2067 steps (43)
[11:19:06] Writing local files
[11:39:17] Completed 910 out of 2067 steps (44)
[11:39:17] Writing local files
[11:59:28] Completed 931 out of 2067 steps (45)
[11:59:28] Writing local files
[12:18:41] Completed 951 out of 2067 steps (46)

[TollhouseFrank]

wow man... nearly a minute difference per frame if i'm not mistaken!

[stratcatprowlin]

wow man... nearly a minute difference per frame if i'm not mistaken!

Look again. Try 5. :)

[TollhouseFrank]

hmm.... man... yer right... my math sucks

[stratcatprowlin]

It's just unreal.Games don't even get that big of a difference with memory differences like these.

[Silver]

I would not mistake amount with dual channel. I am getting 12:32 per frame on 2 x 256 running 250 fsb at 2,2,2,6 1:1 on the 2.8c. That is of course on QMDs.

[stratcatprowlin]

I would not mistake amount with dual channel.

Yes I did take that into consideration.Thanks for posting that if I didn't stress that part.

[ChasR]

The difference between 19.8 minutes per frame and 17.0 minutes per frame is 2 minutes 48 seconds, I beleive. :D

The difference would be greater if the checkpoints on the first rig were set longer than the frame time. On QMDs that virtually useless checkpoint costs 46 seconds per frame. I say virtually useless since mostly the same data is written to disk again only 2 minutes later.

[stratcatprowlin]

The difference between 19.8 minutes per frame and 17.0 minutes per frame is 2 minutes 48 seconds, I beleive. :D

The difference would be greater if the checkpoints on the first rig were set longer than the frame time. On QMDs that virtually useless checkpoint costs 46 seconds per frame. I say virtually useless since mostly the same data is written to disk again only 2 minutes later.

Ok my math sucks too lol. Any how theyre the same rig with swapped memory and checkpoints set to 30.

[ChasR]

The log indicates a checkpoint 15 minutes after the completion of each frame in the first set but not in the second set.

[stratcatprowlin]

Oh maybe I changed it afterwards.