![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
8P with Xeon E7-8870
- Thread starter -alias-
- Start date
so 80 cores with 160 threads at minimum 2.4ghz (probably 2.6ghz turbo??) would come out to about....
eleventy billion ppd
but really, don't think i've seen anyone with 8p yet, 4p will get over 1 million ppd on the better WUs.
get it and let us know the PPD
eleventy billion ppd
but really, don't think i've seen anyone with 8p yet, 4p will get over 1 million ppd on the better WUs.
get it and let us know the PPD
so 80 cores with 160 threads at minimum 2.4ghz (probably 2.6ghz turbo??) would come out to about....
eleventy billion ppd
but really, don't think i've seen anyone with 8p yet, 4p will get over 1 million ppd on the better WUs.
get it and let us know the PPD
Tear has/had an 8P F Socket...
As far as the OP. No idea but interested to see!
Tear has/had an 8P F Socket...
you and your technicalities. i meant a recent gen 8p of this caliber.
I think Patriot has a 4P E7 machine we can extrapolate from. E7's are older that E5's and are less efficient. If memory servers, 10 E7 cores roughly equals 8 E5 cores clock for clock. So as a rough guess assuming 4 x E5 8 cores @ 3.175GHz = 6:40/frame on an 8102:
8:22 for 4 procs
4:11 for 8 procs
~2M ppd on the best units.
This assumes the client scales to -smp 160 and that my original assumption of E7 performance is correct.
8:22 for 4 procs
4:11 for 8 procs
~2M ppd on the best units.
This assumes the client scales to -smp 160 and that my original assumption of E7 performance is correct.
terrabyte89
Weaksauce
- Joined
- Nov 6, 2004
- Messages
- 120
I thought they might have scaled even higher, taking into account a few things. I've got a pair or E5-2680's which are pretty much equivalent to the e5-4650's that you would see in a 4P xeon machine.
As far as I've read the 4p 4650's are right for nearly 1 million PPD on the good work units. My pair of 2680 doesn't make half of their output, in fact they're only good for about 360k on the best WU's. Which seems to be that double the cores/cpu's equals more than a 2 fold PPD increase. Possibly getting closer to 3x.
By that logic I would expect the 8P do more than double the equivalent 4P production.
But as Musky has said the E7's are older tech and running at a lower clock speed than the 4P e5 xeons so they're not really just a doubling of an E5-4650 4P machine.
I think it would be fantastic to see what they could do though, I reckon they might pull out a few surprises. It would easily be the highest PPD in a single machine I've ever seen either way though!
As far as I've read the 4p 4650's are right for nearly 1 million PPD on the good work units. My pair of 2680 doesn't make half of their output, in fact they're only good for about 360k on the best WU's. Which seems to be that double the cores/cpu's equals more than a 2 fold PPD increase. Possibly getting closer to 3x.
By that logic I would expect the 8P do more than double the equivalent 4P production.
But as Musky has said the E7's are older tech and running at a lower clock speed than the 4P e5 xeons so they're not really just a doubling of an E5-4650 4P machine.
I think it would be fantastic to see what they could do though, I reckon they might pull out a few surprises
. It would easily be the highest PPD in a single machine I've ever seen either way though!
D
Deleted member 12106
Guest
You'll also see bonus points for finishing sooner on 4p over 2p.
402blownstroker
[H]ard|DCer of the Month - Nov. 2012
- Joined
- Jan 5, 2006
- Messages
- 3,242
I am pretty sure there has been on evaluation an 8P Xeon setup and its scalability falls on it face. 2 4P Xeon machines bets it PPD wise. Too lazy to do the search myself, but it is on this site.
Mike_Shaffer
n00b
- Joined
- Sep 13, 2013
- Messages
- 38
Could it run two clients with half the cores locked to each client?
I think it can, but I'm pretty sure it will not make any difference in PPD.
PPD would be worse, the bonus is exponential. Halving the tpf more than doubles the PPD
Dead Things
Limp Gawd
- Joined
- Jul 8, 2011
- Messages
- 152
From a PPD/W perspective, LGA1567 leaves a lot to be desired. My 4P 8870's did about 680k PPD on P8102 while pulling 850-900W from the wall, IIRC. Hopefully you can use that to help estimate tpf.
I was afraid that the power consumption will be over 12 - 1300W for a P8 like this. But if we double what you suggest we end up at about 16 - 1700W and it becomes too much for my finances, that's for sure. I already use about 2500W with the rigs I have.
I'm still excited about the figures tear will come up with, so I wait a bit before I decide.
As a point of reference, my 4P E7-4860 (80 threads) does the following:
8101 TPF 12:24avg, 11:59min, 397-418K PPD
8102/03 TPF 9:11avg, 624K PPD
8104 TPF 6:24avg, 709K PPD
8105 TPF 9:0avg, 614K PPD
Dead Things
Limp Gawd
- Joined
- Jul 8, 2011
- Messages
- 152
I was just going off memory, but your results seemed lower than I remembered, so it inspired me to look up my old HFM logs...
P8101: Avg. Time / Frame : 00:11:33 - 442,720 PPD
P8102: Avg. Time / Frame : 00:08:52 - 657,987 PPD
P8103: Avg. Time / Frame : 00:08:58 - 647,047 PPD
P8104: Avg. Time / Frame : 00:06:18 - 724,553 PPD
...Sorry, I overstated the P8102 performance in my previous post.
My guess was very close then. Using this as a base line, that is 8:40/frame to get 680K ppd. Half that is 4:20, which comes out to 193K ppd.
And that gets us to 186K ppd on an 8P version.
Dead Things
Limp Gawd
- Joined
- Jul 8, 2011
- Messages
- 152
Think you mean 1.86M PPD. But would scaling from 80 to 160 threads be 100% efficient like that?
I assume you are joking! In my head PPD will roughly double from a 4P rig to a 8P with the same chips.
and by 186K ppd, I of course meant 1.86M ppd....
It is close to half the frame time going from 4p to 8p, which is very close to triple the ppd.
and by 186K ppd, I of course meant 1.86M ppd....
It is close to half the frame time going from 4p to 8p, which is very close to triple the ppd.
It is really impressive, only sorry that the power consumption follows a similar progression.
Mike_Shaffer
n00b
- Joined
- Sep 13, 2013
- Messages
- 38
I only meant that if it didn't scale well from 4P to 8P then it could act as 2 x 4P. Just curious, since there were some other posts about power draw, what about size? I mean, unless the processors are on riser cards, this has to take up some serious
real estate.Sincerely,
Hi, are these PPDs produced by a 4P E7-8860 or by a 4P E7-8870? In an old post you mentioned it was a 4P E7-8860 and the PPD was nearly the same.
Last edited:
Hi -alias-, my last analyzation about this question is here:
http://hardforum.com/showthread.php?t=1716471&page=3
One whole year is past, hope it's still helpful.
my 1567 4p 10c 2.4ghz does 2/3rds the output of the 2.7ghz 2011 octo 4p
I will collect TPF from it in the morning.
Hi -alias-, my last analyzation about this question is here:
http://hardforum.com/showthread.php?t=1716471&page=3
One whole year is past, hope it's still helpful.
Thanks quickz.
It is very helpful, and I am sorry that I did not remember that I have asked about this before. Back in 2005 I had 2 brain stroke that destroyed something so I am still struggling a bit to remember from the recent past. But now that I read what I / we wrote a year ago, it comes back to me.
Dead Things
Limp Gawd
- Joined
- Jul 8, 2011
- Messages
- 152
Good point. I think it might be a bit of both since it was the same machine that got upgraded. The 8860's might be contributing to some of those numbers for the older units. I cannot say for certain, though - I got rid of that rig a number of months ago.Hi, are these PPDs produced by a 4P E7-8860 or by a 4P E7-8870? In an old post you mentioned it was a 4P E7-8860 and the PPD was nearly the same.
http://forums.evga.com/FindPost/1094557
Unfortunately, those 8P numbers from sfield are estimated. Last data on folding forum had bigadv scaling linearly to -smp 64 (i.e. almost 100%) and starts to drop off to to about 90% on a gentle curve up through smp -128 (depending on the WU, bigger WUs scale better). Above 128 the scaling takes a big hit. Have not seen numbers above 160 yet.
PPD would probably be better running 2 clients, one at -smp 128 and one at -smp 32.
Unfortunately, those 8P numbers from sfield are estimated. Last data on folding forum had bigadv scaling linearly to -smp 64 (i.e. almost 100%) and starts to drop off to to about 90% on a gentle curve up through smp -128 (depending on the WU, bigger WUs scale better). Above 128 the scaling takes a big hit. Have not seen numbers above 160 yet.
PPD would probably be better running 2 clients, one at -smp 128 and one at -smp 32.
Dead Things
Limp Gawd
- Joined
- Jul 8, 2011
- Messages
- 152
Can you even get Windows to do anything more than -smp 64? I thought with processor groups, 64 logical threads was the most Windows could deploy to any single task...
Dead Things
Limp Gawd
- Joined
- Jul 8, 2011
- Messages
- 152
Yeah, but isn't it weird to say scaling from 64 threads to 128 threads is 90% efficient when it's actually impossible? I realize that's not what 7im was saying, but it is what tear says 7im was describing...
If interested:
Concepts on systems with more than 64 procs
Limitations on current versions
To my understanding, there is no 64-bit version of FAHcore available on Windows, only on Linux. All 32bit Win apps are limited to 32 cores anyway.
I dont see anything in that post relating to windows? It is entirely possible in linux. Then again I am most likely mis-reading something