memory bandwidth effect on folding

[Patriot]

Hey guys... I have been tweaking my amd setups for ram bandwidth... trying to eek every last second out of my tpf....

numa off or numa on... I honestly have not seen much difference...

ram channels on the other hand... have had a huge difference... dropped 3min on my 24core magny cours...

and I noticed that the less ram had little to no effect on the c32 system I stole from to boost the g34 box... because it only does dual channel....

I noticed also that the lesser of the three c32s dual 2.1 hexa vs dual 2.3 hexa was getting 1:20 faster tpf than one of the 2.3s on the same wu...

difference.... single vs dual channel... yeah I turned them off lol... swapped ram around...


I will have to work on getting all the intel boxes into tri-channel mode.... or does it not help them as much?

 

[Tobit]

ram channels on the other hand... have had a huge difference... dropped 3min on my 24core magny cours...

Yeah, some of us has seen this before although no in-depth study has ever been done.

 

[Patriot]

Yeah, some of us has seen this before although no in-depth study has ever been done.

not leaving good enough alone seems to be having evil payback...
that one server doesn't seem to like dual channel mode...

I think I should add this to my sig ...

"have you killed a proto today... I have it was fun!"

well... all the movement freed up some ram lol

so... down one dual c32 rig @ 2.3 slowest one of those... tpf 28:30 trying to speed it up and its cranky now...
added single g34 12cores @ 2.2 quad channeled it....

somehow I think it will even out...

on my todo list is to get tri channel working for the 2 dual intel hexacore systems....
turn another server from single g34 8 core to dual @ 2.4 ....
add dual intel 8core....
...and fix the mad proto....

add that to the end of do my real work....

 

[Kendrak]

Yeah, some of us has seen this before although no in-depth study has ever been done.

He lives!

On the i7 systems when they first came out, one of the main reasons for going with a 1366 board was the tri memory chann. There was a big enough difference that it was worth the added cost.

 

[Patriot]

He lives!

On the i7 systems when they first came out, one of the main reasons for going with a 1366 board was the tri memory chann. There was a big enough difference that it was worth the added cost.

I have seen the benches for day to day stuff.... tri don't matter for that actually...pretty much the same for games...

but it seems f@h is bandwidth hungry...

 

[sirmonkey1985]

I have seen the benches for day to day stuff.... tri don't matter for that actually...pretty much the same for games...

but it seems f@h is bandwidth hungry...

the entire WU is run directly off the memory, so memory bandwidth plays a key factor in F@H. its also why the G34 boards do so well on F@H with the lower clocked magny-cours compared to the xeons.

 

[Patriot]

the entire WU is run directly off the memory, so memory bandwidth plays a key factor in F@H. its also why the G34 boards do so well on F@H with the lower clocked magny-cours compared to the xeons.

and I would assume that the more cores the more bandwidth or lack of would be noticeable...

I think my dual quads with no ht are gonna get robbed and make sure all my dual hexacores have tri-channel running...

 

[sirmonkey1985]

more cores = more bandwidth required since theres more data being changed between threads.. or so i would think.

 

[Mr. Pedantic]

Not to be pedantic or anything D), but the work is eke, and it is spelled with the k in the middle.

the entire WU is run directly off the memory, so memory bandwidth plays a key factor in F@H. its also why the G34 boards do so well on F@H with the lower clocked magny-cours compared to the xeons.

I thought Nehalem-EX already had support for quad-channel. Or does nobody have one? In any case, I think LGA2011 will have quad-channel support; 1 DIMM per channel, is what I heard. Be nice to see how that bandwidth increases ppd.

 

[Patriot]

Not to be pedantic or anything D), but the work is eke, and it is spelled with the k in the middle.


I thought Nehalem-EX already had support for quad-channel. Or does nobody have one? In any case, I think LGA2011 will have quad-channel support; 1 DIMM per channel, is what I heard. Be nice to see how that bandwidth increases ppd.

nehalem-ex is 1567 ... chips are 2k a pop... and the board is more than that... if you can find one... no i don't think anyone uses them...

LGA 2011 has pcie 3 and quad channel... and I can't tell you how not ready it is....

 

[APOLLO]

Tobit, where the heck have you been?!?

I thought Nehalem-EX already had support for quad-channel. Or does nobody have one? In any case, I think LGA2011 will have quad-channel support; 1 DIMM per channel, is what I heard. Be nice to see how that bandwidth increases ppd.

Pedantic, nice to see you posting. you're right about the memory sub-system being quad channel on Nehalem-EX. I have yet to come across anyone with a Beckton post over here and doubt someone has ready access to such a system. However, there was an individual last year who posted on the FF with a 64-thread (?) system but A2 -bigadv scores and they were disappointing in light of the thread count.

LGA 2011 has pcie 3 and quad channel... and I can't tell you how not ready it is....

That isn't encouraging news from the sound of it.

 

[Patriot]

Tobit, where the heck have you been?!?

Pedantic, nice to see you posting. you're right about the memory sub-system being quad channel on Nehalem-EX. I have yet to come across anyone with a Beckton post over here and doubt someone has ready access to such a system. However, there was an individual last year who posted on the FF with a 64-thread (?) system but A2 -bigadv scores and they were disappointing in light of the thread count.

That isn't encouraging news from the sound of it.

Its not due out for sometime according to the intel released time schedules...

 

[Zink]

For single socket stuff like Sandy and BD I would assume because there are only two channels that buying cheap 2133mhz ram would only make sense. I really want to see someone post bigadv numbers with their ram at full 2000+ clocks then take the exact same WU and system and put the ram at 1333 or 1600mhz. BD is still only going to be running 8 threads of F@H so the effects of ram bandwidth probably won't change too much from i7.

 

[APOLLO]

Its not due out for sometime according to the intel released time schedules...

/true dat

 

[sirmonkey1985]

For single socket stuff like Sandy and BD I would assume because there are only two channels that buying cheap 2133mhz ram would only make sense. I really want to see someone post bigadv numbers with their ram at full 2000+ clocks then take the exact same WU and system and put the ram at 1333 or 1600mhz. BD is still only going to be running 8 threads of F@H so the effects of ram bandwidth probably won't change too much from i7.

they have been posted before. i'll find them..

http://hardforum.com/showpost.php?p=1036648028&postcount=19


http://hardforum.com/showpost.php?p=1036651859&postcount=32

those were done with the benchmark application musky made so they are identical..

2133 vs 1600 is a difference of about 1 minute on the TPF for a P6900 WU or about 4k PPD.

 

[Toconator]

2133 vs 1600 is a difference of about 1 minute on the TPF for a P6900 WU or about 4k PPD.

Interesting, that. I have been downclocking my RAM in rig #3 to 1333 since taking a HD4850 out of it and using the 4290 IGP. Seems the sideport runs at 1333 and it was crashing all the time with the RAM @ 1600, setting them both to 1333 has stopped the memory related crashes (but now I lost a core, only 3 showing up on a 1 month old 965BE). Been waiting on a deal for another good GPU folder and will definitely crank the RAM back to 1600 when I do so.

 

[Patriot]

Interesting, that. I have been downclocking my RAM in rig #3 to 1333 since taking a HD4850 out of it and using the 4290 IGP. Seems the sideport runs at 1333 and it was crashing all the time with the RAM @ 1600, setting them both to 1333 has stopped the memory related crashes (but now I lost a core, only 3 showing up on a 1 month old 965BE). Been waiting on a deal for another good GPU folder and will definitely crank the RAM back to 1600 when I do so.

I have found a core to dissappear from time to time if unlocker is set on auto... just put it in manual and leave all cores enabled...

 

[sirmonkey1985]

Interesting, that. I have been downclocking my RAM in rig #3 to 1333 since taking a HD4850 out of it and using the 4290 IGP. Seems the sideport runs at 1333 and it was crashing all the time with the RAM @ 1600, setting them both to 1333 has stopped the memory related crashes (but now I lost a core, only 3 showing up on a 1 month old 965BE). Been waiting on a deal for another good GPU folder and will definitely crank the RAM back to 1600 when I do so.

sounds like a problem with the board and not the CPU. wonder if something in the bios settings like the "unlocker" is causing that core to disapear. the other thing you can check is the boot.ini advanced settings msconfig/boot/advanced settings.

 

[Toconator]

I have found a core to dissappear from time to time if unlocker is set on auto... just put it in manual and leave all cores enabled...

Ahh, thought of that already and no diff. Apparently was BSOD'ing frequently and then stopped. Of course this rig is in the family room and is used by the wife and kids so no one said anything for a few days. That's when I noticed 1 core offline. Unlocker is on manual and the 4th core is set to on in BIOS but it's not there after boot. Still BSOD's once in a while and it's reset the mobo back to defaults once, so I suspect it came up with an error on boot and someone pressed set to defaults to get it to run. Then I ran into the RAM speed issue again and had to reset to 1333 and all relevant settings to manual. I'll have to RMA it and get a cheap CPU to fill in for the meantime so the wife can get her email and effing facebook. It only smp folds with no GPU so it will be a small hit until it's fixed.

 

[sirmonkey1985]

try my suggestion with the msconfig. I've seen weird things where windows will randomly change the core count to 2 or 3 cores which will cause the 4th core to completely disappear in windows.