![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Top 5 PPD/Watt Rigs in the [H]orde
- Thread starter Kendrak
- Start date
My 4p Magny Cours 6172 @ 2.1 GHz turns in 15:15 frame times on a 6903 for 445K ppd. Don't have current power draw info though. My Kill a Watt is being used for another project.
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
2P x5570 @ 3.2Ghz
310W at the wall
103K PPD on 6903
00:40:23 TPF 6903
333 PPD/W
310W at the wall
103K PPD on 6903
00:40:23 TPF 6903
333 PPD/W
Last edited:
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
Its amazing how the PPD/Watt has jumped up in the past year.
R-Type
[H]ard|DCer of the Month - October 2011
- Joined
- Mar 6, 2006
- Messages
- 2,801
2P G34 Extra Sexy Dodecas @3.2ghz/2.4ghz NB
286k PPD (6904 avg)
520W
550 ppd/w
286k PPD (6904 avg)
520W
550 ppd/w
R-Type
[H]ard|DCer of the Month - October 2011
- Joined
- Mar 6, 2006
- Messages
- 2,801
You should be seeing higher ppd than that with the 3.1ghz turbo folding frequency on those hexes. Make sure you have numa enabled and linux installed so you can get the 6903/6904 units, you should be closer to 130k ppd I would think.
Everything is setup for 6903/4 but I hvae never received one on either box, this is using v7 client. I'm going to reinstall everything put and v6 on the 5640 box and see what that does.
Sgraffite
Supreme [H]ardness
- Joined
- Jan 10, 2006
- Messages
- 4,405
It really is, what actually changed so much that people are able to get 3 to 4x the PPD/Watt? It seems both Intel and AMD CPUs are able to hit over 800 PPD/Watt now. I'm not the most experienced, but I know Sandy Bridge brought some power saving to the table for Intel, but I'm unsure of what AMD has done that has changed in the last year.
The PPD/Watt change is much more than I expected since 1 year ago.
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
A lot has to do with the -bigbeta units since then which I get 333 PPD/W. Even on the old bigadv 6900 I still get 200 PPD/W and regular SMP gets me about 129 PPD/W ish. So I really think its a combination of work unit changes, and we may have just become more efficient. There are a lot of tweaks and optimisations that we use such as running native linux with the kraken affinity wrapper. I was not folding a year ago but its still amazing how much the PPD/W has changed even with the same hardware. Even on the same work units with the same hardware we get more points now and it just goes to show how much more effort we put in to maximizing our PPD and lowering power consumption.
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
Those are very nice chips he has, they have 3.33 Ghz Turbo! (25x Muliplier)
It does seem to be that v7 does not get -bigbeta. Those hexes should be destroying some PPD... My dual x5570 quad cores at 3.2Ghz get 103K PPD on 6903 just for reference. Those X5670's have a 133mhz clock speed advantage and 2 extra cores! They are 4% faster clocked and have 50% more cores so 56% more processing power. I would guess 00:25:52 TPF on 6903 under good conditions which would net you about 201K PPD if my numbers are correct.
Hey I'll take regular -bigadv, it's processed nothing but A3 SMP for the last 3 days!!
At the time we originally did this, you needed Windows to run bigadv (or Wine in Linux), the SR-2 had just came out (jebo, Kendrak, and Vaulter had the first 3 for the Horde), and hex Xeons and 12-core G34s were well over $1,000 per chip. Some of the big events:
- MIBW showed up with his herd of dual hex SR-2 machines that would have beaten any of the original boxes listed, and did a huge portion of the initial overclocking work for SR-2s.
- The L5640 "extra spicey" chips followed soon afterwards, making dual hex SR-2 rigs a more practical possibility.
- 10e (and I am sure others) figured out how big of an effect NUMA settings in multi-proc boards has
- bigadv came to Linux, and 10e and others showed us what was really possible with the G34 platform with Linux and bigadv.
- We discovered the benefits of the BFS scheduler for 24 or fewer cores
- tear developered The Kraken wrapper - given pseudo NUMA support for the FAH core.
- Stanford released the 6903 and 6904 work units.
- The availability of G34 chips outside of retail outlets increased, making them available to more of us.
These were all big events that have happened in the last year+, and the sum total has lead us to where we are now with the big 48 core machines approaching 900 ppd/W and several dual proc options giving 400+ ppd/W. It has definitely been a fun year!
Last edited:
R-Type
[H]ard|DCer of the Month - October 2011
- Joined
- Mar 6, 2006
- Messages
- 2,801
Yes, they have been running that way for about a month. I will be stepping up to 3.3ghz shortly due to the advent of cooler weather, I'd love to break 300k ppd with it.
This is the system:
http://hardforum.com/showthread.php?t=1639667
Sgraffite
Supreme [H]ardness
- Joined
- Jan 10, 2006
- Messages
- 4,405
Awesome! That is very informative for people like me who don't pay enough attention to keep up with all the changes
BTW good job on your unbuntu folding guide, well organized and very easy to follow. I'm a big fan of putting linux commands in the code tags for readability.
When I have a bum units and am pulling 600k ppd with my 4p... at 915w thats still 655ppd/w non overclocked 4ps typically rank in the 800s ppd/w
R-Type
[H]ard|DCer of the Month - October 2011
- Joined
- Mar 6, 2006
- Messages
- 2,801
That definitely beats SR-2 ppd/w, seems Intel's engineers were right in assigning such low clocks to the server parts.
AMD too. Patriots ppd/w goes to hell when he overclocks too.That definitely beats SR-2 ppd/w, seems Intel's engineers were right in assigning such low clocks to the server parts.
that antec PSU is a POS, when i can afford to swap it out i'll plug the killowatt back in and see what its pulling then.
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
What clocks are these guys running at?
The L5640 looks a little dissappointing (PPD wise not PPD/W
), and the X5670 seems about right.My dual X5570's get 104K PPD on 6903
8 x 3.2Ghz = 25.6Ghz, 104K PPD
Your L5640's get 112K PPD
12 x 2.8Ghz = 33.6Ghz, 112K PPD
33.6/25.6 = 1.3125, so your chips have a 31.25% performance advantage over mine.
112,000/104,000 = 1.0769 (rounded down) which is only 7.69% more PPD
Now 12 x 2.26Ghz (Stock L5640 no turbo) = 27.12Ghz
27.12/25.6 = 1.059375 which is 5.9375% faster than my quads
So factoring for the exponential curve, and upload time your L5640 hexes are running at stock! no turbo!
EDIT: The X5670's are 12 x 3.33Ghz = 39.96Ghz, 39.96/25.6 = 1.5609375, 56.09375% faster than my quads.
1.5609375 x 104,000PPD = 162,337.5 PPD
Not sure what your upload times are but it seems the bonus curve is less pronounced here. Are you using HFM estimated points? Thats what I am using for my calculations.
Last edited:
the L5640's are at 2.26 - not sure if they turbo or not, if they do i think its only to 2.4. The x5670's have turbo and they go to 3.2. Since they are both on linux and i hate linux more than jebo does - i'll wait for my new daily driver to arrive and load windows and see what CPU-z is saying about spicy L5640's
Quick maths shows that the L5640's do not turbo but Musky would know as they were his originally.
If they do turbo then the rig is running slow
Quick maths shows that the L5640's do not turbo but Musky would know as they were his originally.
If they do turbo then the rig is running slow
R-Type
[H]ard|DCer of the Month - October 2011
- Joined
- Mar 6, 2006
- Messages
- 2,801
The spicy chips should be turboing 1x versus the 2x for retail, I haven't seen any that do this differently.
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
They do turbo just fine but it's only to 2.6 GHz (x1) when all cores are fully loaded.
That is what I've seen.
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
Ok, my maths was a little off, but I was in the ballpark
If they are turboing then something isn't running right. By my maths
112000/24/2.4 = 1944 PPD per core per ghz. Mulitply that out by the x5670 numbers and I don't get the 165k that the 5670's are currently reporting.
Start with the 165k and do the maths and i should be getting 123k from the L5640's.
I'll have a look tomorrow and see if i can see anything. Is there any similar utility to CU-z on linux that will tell me what everything is currently running at?
112000/24/2.4 = 1944 PPD per core per ghz. Mulitply that out by the x5670 numbers and I don't get the 165k that the 5670's are currently reporting.
Start with the 165k and do the maths and i should be getting 123k from the L5640's.
I'll have a look tomorrow and see if i can see anything. Is there any similar utility to CU-z on linux that will tell me what everything is currently running at?
Untitledone
[H]ard|DCer of the Month - April 2012
- Joined
- Feb 28, 2011
- Messages
- 1,486
I dont think they are Turboing, the numbers work out to them being at stock.
R-Type
[H]ard|DCer of the Month - October 2011
- Joined
- Mar 6, 2006
- Messages
- 2,801
PPD is a non linear function of compute power while your direct comparison computational approach is linear. Your reasoning would work when trying to predict one tpf given another but cannot be applied to ppd. The fact that you get a bigger point increase for added speed than a linear comparison would suggest is Standford's way of rewarding people for upgrading hardware and the crux of the bigadv controversy of the last couple years.
Your rigs are running fine.