• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

New single boxen PPD record

Wow.....

Hey Tiger.... found out you next step on your upgrade path.
 
That user is running the Xeon X7560 ( Nehalem-EX) which costs $3,692 per processor.

It has 8 cores / 16 threads on each cpu.

That system costs $14,700 just for the processors.
When you add in the motherboard, ram, and other components, that is a very expensive system.

Not the best points per dollar in any situation.
 
EvilAlchemist said:
That user is running the Xeon X7560 ( Nehalem-EX) which costs $3,692 per processor.

It has 8 cores / 16 threads on each cpu.

That system costs $14,700 just for the processors.
When you add in the motherboard, ram, and other components, that is a very expensive system.

Not the best points per dollar in any situation.
Click to expand...
Definitely not. This guy is running 64 threads with Nehalem-class processors and 'only' producing 114k PPD?? What does this tell us about a SR-2 with dual hex-cores and 24 threads...?? :confused:
 
APOLLO said:
Definitely not. This guy is running 64 threads with Nehalem-class processors and 'only' producing 114k PPD?? What does this tell us about a SR-2 with dual hex-cores and 24 threads...?? :confused:
Click to expand...

Could it have a lot to do with clock speed? - the same thing was noted with the new Magny-cours opterons, my view is that the same or fewer cores at a higher clock is better, Slash2 proved that on FF when he went from 1.9 to 2.3ghz opterons. If we could get the xeons upto 2.6-2.8 we might see something
 
Part of the issue is he was running the SMP -bigadv.
We all know there are some issues when running and even number of nodes.
(SMP 15 vs SMP 16)

Going all the way up to 64 nodes would cause a lot of issues i am sure.

I bet it would have better performance on the SMP2 A3 cores.
 
That and there is a diminishing return when you reach a certain # of cores.

Yes they are designed for parallel computing, but only up to a certain point.

We have seen this on some GPU WU's lately.

I would expect an SR-2 with OC'd 5630's to perform quite similarly from a PPD perspective. And thats with only 24 threads
 
Would prob get better PPD from running two instances of BigAdv.
 
EvilAlchemist said:
Part of the issue is he was running the SMP -bigadv.
We all know there are some issues when running and even number of nodes.
(SMP 15 vs SMP 16)

Going all the way up to 64 nodes would cause a lot of issues i am sure.

I bet it would have better performance on the SMP2 A3 cores.
Click to expand...

I'd like to see A3 numbers on that system as well.

But good point on a costly box "only" cranking 114k ppd.
 
Nathan_P said:
Could it have a lot to do with clock speed? - the same thing was noted with the new Magny-cours opterons, my view is that the same or fewer cores at a higher clock is better, Slash2 proved that on FF when he went from 1.9 to 2.3ghz opterons. If we could get the xeons upto 2.6-2.8 we might see something
Click to expand...
Even at 2.0GHz that's a low score IMV, and I'm almost convinced the clockspeed was higher than that.

EvilAlchemist said:
Part of the issue is he was running the SMP -bigadv.
We all know there are some issues when running and even number of nodes.
(SMP 15 vs SMP 16)

Going all the way up to 64 nodes would cause a lot of issues i am sure.

I bet it would have better performance on the SMP2 A3 cores.
Click to expand...
I remember you posting about that a while ago. To me, it seems there is a scaling inefficiency above a certain number of threads that is creating a ceiling in performance. If true, maybe he should run 4 -bigadv clients with 16 threads each?

/shrug
 
APOLLO said:
Even at 2.0GHz that's a low score IMV, and I'm almost convinced the clockspeed was higher than that.

I remember you posting about that a while ago. To me, it seems there is a scaling inefficiency above a certain number of threads that is creating a ceiling in performance. If true, maybe he should run 4 -bigadv clients with 16 threads each?

/shrug
Click to expand...

Even if he was running 4 -bigadv clients with 16 threads each, ppd would only be around 30k PpD per client, or 120k PpD for the box. So, if there is some inefficiency, he's well within the ballpark figure, especially without knowing the clock speed on each core.

EDIT: I think the clock speed for the X7560 id 2.26Ghz, or the same as the x5520. Which makes his 114k PpD well within reason.

On another note, 980x @ 4.0Ghz is doing 42k PpD, which is pretty tasty :D
 
Last edited:
crazjayz said:
Even if he was running 4 -bigadv clients with 16 threads each, ppd would only be around 30k PpD per client, or 120k PpD for the box. So, if there is some inefficiency, he's well within the ballpark figure, especially without knowing the clock speed on each core.
Click to expand...
Well, what you're saying would be 100% true if there was no bonus calculation since the resultant PPD would show a near linear scale, but that isn't the case with -bigadv. Taking the bonus scheme into account actually does reveal inefficiencies. There are at least three possible scaling effects to consider: F@H client scaling, OS thread scaling efficiency and of course, microarchitecture scaling efficiency (hardware). There are no systems in my knowledge whether software or hardware-based that demonstrate 100% efficiency in regards to high order scaling, and this becomes more apparent when there are several scaling effects to consider with extremely high thread counts as is the case with this kind of application.

On another note, 980x @ 4.0Ghz is doing 42k PpD, which is pretty tasty :D
Click to expand...
Precisely the crux at mine and Evil's comments. For the stellar price tag, you could probably net ~500k PPD if not higher by purchasing a farm of a dozen or more i7 boxen. That is at least 5 times more PPD, which would place anyone easily within the top 5 PPD single producers in the entire world.
 
If you're space limited, then a dual processor box may be the way to go, but the cost of entry is high.

As you've all said, if you have the space, a bunch of i7 920/930 boxen will net you greater PpD with a lower initial cost.

The 980x may get you 42K PpD, but at a cost of $1000 per CPU. For that price (assuming a nearby MC), you can pick up nearly 2 complete i7 930's (including mobo's, ram, psu and all the fixin's) for less and get yourself at least 50K Ppd. Probably closer to 55K.

Makes me wonder about the frugality of the SR-2.
 
A couple of hours latter a guy with 48 cores of hex opterons @ 2Ghz comes along with 9min 30s frames @ 160k ppd. Looks like he was only running -smp 45. These systems don't make sense at all for 100-200k ppd home farms.
 
APOLLO said:
Well, what you're saying would be 100% true if there was no bonus calculation since the resultant PPD would show a near linear scale, but that isn't the case with -bigadv. Taking the bonus scheme into account actually does reveal inefficiencies. There are at least three possible scaling effects to consider: F@H client scaling, OS thread scaling efficiency and of course, microarchitecture scaling efficiency (hardware). There are no systems in my knowledge whether software or hardware-based that demonstrate 100% efficiency in regards to high order scaling, and this becomes more apparent when there are several scaling effects to consider with extremely high thread counts as is the case with this kind of application.

Precisely the crux at mine and Evil's comments. For the stellar price tag, you could probably net ~500k PPD if not higher by purchasing a farm of a dozen or more i7 boxen. That is at least 5 times more PPD, which would place anyone easily within the top 5 PPD single producers in the entire world.
Click to expand...

Ah, I see your point about scaling now. My bad before on not understanding that. I knew that -bigadv scaled, but for some reason my mind blanked (possibly due to the number of core he's folding with or something :D). Yeah, assuming dual 5520s have a TPF of 28 min, he should be getting 4 times better than that, or 7 min TPF. Which in that case would net him almost 260k PpD.

And I agree, a bunch of 980s would be much better compared to the +3k cost of the CPUs, each.
 
The system also used 1.13kW with 256GB of ram which is about i7 920 level, fail for 20 grand though.
 
Zink, Do you have a link to that 160k ppd rig? If my math is right you could build a 4-way opteron 6168 rig with 32 GB RAM for under $6000(cheap, huh) that should get 150k ppd with around 600W of power.
 
Last edited:
Don't forget clock speed plays a role here too

I may be way off, but this made sense In my head
and don't bite my head off for cores, I know what they really are, just role with me here


64 cores * 2.26 ghz = about 144
12 cores * 4.0 ghz = 48
48 is exatcly 1/3 of 144, so in theory the uber rig should be 3 times faster
not acounting for the bonus even, the uber rig does not hit that mark, and falls way short if we could adjust for bonus

This shows we have diminishing returns (or in other words an inefficent multi threaded app) when you compare core speed to core population
so having fewer cores at higher speeds is better, we just need to find the highest we can go before returns diminish

this is proved on the other side of the scale when you compare hex to quad
with 50% more cores, the same oc on a hex will net you more than 50% (from 26k to 42k)
so we can see, at the lower ends, adding more cores at a fast speed pays off, but we need to find the point where adding cores becomes a loss

It's just a feeling but I don't think we will find that with our limit with SR-2
worst case scenario, all we have to beat is 50k, and that shouldn't be too hard

another thing we need to consider is multi socket scaling. Maybe the hex kicks ass compared to this uber rig because folding doesn't scale well over multiple sockets, we have to consider that too
Posted via [H] Mobile Device
 
3 posts down from the xeon rig in the op link. username Punchy. It looks like he either has another rig with even more cores or he was experimenting with using more threads than he has cores, I'm not sure.
 
Zink said:
A couple of hours latter a guy with 48 cores of hex opterons @ 2Ghz comes along with 9min 30s frames @ 160k ppd. Looks like he was only running -smp 45. These systems don't make sense at all for 100-200k ppd home farms.
Click to expand...

This also says something else, 48 real cores on the older istanbul architecture is better than 32+32HT on the newer nehelem processors. Several things are evident here IMHO

1. Clock speed matters - double the cores of the 101k system but at a lower clock does work out to be a near perfect progression (PPD/# of cores/clock = PPD per Ghz per core).Work that backwards and you get 144k so you can predict ppd for a given cpu and no. of cores according to the different speeds available and as proven does so quite well.
(101/24/2.8) =1503PPD/Ghz (1503*2.0*48=144kPPD)
2. For multicpu -bigadv folding AMD cannot be counted out. I'm not a fanboy as at the lower end intel trounces AMD in terms of PPD and PPD/watt
3. HT helps a lot - but real cores help more
4. As pointed out these are uber rigs, beyond the reach of damn near all of us - but they do make for interesting reading, and are very cool:D:cool:

Just as an aside, swapping the 2.0 for 2.8 GHz chips would in theory take the PPD upto 203k:eek:
 
Last edited:
Oh, that is an 8x 6 core rig. That is not a Magny cours, it's an Istanbul system. Magny Cours brings DDR3, lower power, and better socket interconnections to the table so it should outperform that 160k ppd system.
 
Most of us? Even our top spenders can only do 2-3k a year, these rigs are beyond any of us, even us mega folders. Even atlas would be pushed kinda hard for this, this is strictly for those that have no financial limitations, those with backers
Posted via [H] Mobile Device
 
Vaulter98c said:
Most of us? Even our top spenders can only do 2-3k a year, these rigs are beyond any of us, even us mega folders. Even atlas would be pushed kinda hard for this, this is strictly for those that have no financial limitations, those with backers
Posted via [H] Mobile Device
Click to expand...

True, i stand corrected, atlas is eying up fermi cards tho,
 
dual g34 board - $450
dual 8x2ghz Magny-Cours - $600
16 GB ddr3 133 - $ 400

might make an 80k ppd rig that would be more efficient than 3x i7 920@3.8Ghz for a lower price.
350w instead of 750w
cheaper than the 3 i7s after you include cases and power supplies
 
So a dual hex (24 cores counting threads) ties with a 24 real core amd rig, both in the 2.x ghz range

however, SR2 brings the ability to overclock, and it's pretty obvious core count comes in second to core speed (that dude with the q66 that could still do bigadv)

I doubt well see any opteron board that will provide OC abilities that the SR2 brings, the ability to mix and match CPUs (crucial for lowering initial cost) and the ability to use standard ddr3

My money is on intel this round (obviously) but I'm glad AMD is starting to land some good punches
Posted via [H] Mobile Device
 
Vaulter98c said:
however, SR2 brings the ability to overclock, and it's pretty obvious core count comes in second to core speed (that dude with the q66 that could still do bigadv)
Click to expand...
You are 100% right. Frequency trumps core count in efficiency when the frequency is high vs a high core count where the frequency is moderate and scaling problems kick in. That is based on the results I am reading in this thread. I have yet to see what a multiple -bigadv client deployment will garner in these newfangled quad and octal socket systems. Why haven't these guys tested multiple clients when there is an apparent performance ceiling?

I doubt well see any opteron board that will provide OC abilities that the SR2 brings, the ability to mix and match CPUs (crucial for lowering initial cost) and the ability to use standard ddr3
Click to expand...
It won't happen this year that's for sure unless Tyan produces something, but they never have in a decade so doubt they'll do it now. Too bad, it would have been great to see an enthusiast G34 board.

My money is on intel this round (obviously) but I'm glad AMD is starting to land some good punches
Click to expand...
Let's hope there's a lot of pounds per sq in. in those punches otherwise AMD will lose what they have left.
 
Last edited:
I think just the fact that they can get close to intel speaks alot, before this opterons were only looked at for their price, which is partly because the xeons were so good and amd wasn't intel had no reason to lower the price of nehalem. Now they can get close, and while it's not much, it's a damn lot more than they had before. AMD is taking too long on BD, hence the 9% market share in servers, but now they have something to hold the tide until BD gets here.

Now what gets me is that AMD has never really been one for clock speed, and in the past and in that review, it shows, high clock speed is the key with how current apps run, and it will stay that way for some time I think. They pointed out IPC is vital, and AMD has never been one to move on that idea

Plus, intel is in a better stance to cut prices to fight back, AMD is stretched too tight right now. And we have no clue on MSRP of motherboards, ithink it's safe to say anything good amd has will be close in price to SR2, and the OC ability alone will make the value of the intel platform shine. I foget the test setups, but at 60k stock, I'm Thinking 2P systems. 24 cores at 2.2 ghz would be sweet, but with the higher multi and superior build of the SR2, and 4+ghz dual 5650 rig would blow them away, not only in folding but in many applications. Not to mention we can use normal ram and i7 heatsinks, both of which would save money in situations where the results won't cost lives or something. Even a mild OC on ECC ram, like 3.2 or 3.4 ghz, would add crazy value to the equation.

I am hoping that my dual 5530's can OC to a very easy target of 3.8 ghz, where without factoring in the bonus would net me over 50k. With the bonus scaled in, dual last gen CPUs could come really close to the level of stock bleeding edge tech. Since I'll be in windows first I think, I'd be willing to try some of these tests to see where I come in. Any clue where I can get the tests they use?
Posted via [H] Mobile Device
 
Vaulter98c said:
So a dual hex (24 cores counting threads) ties with a 24 real core amd rig, both in the 2.x ghz range

however, SR2 brings the ability to overclock, and it's pretty obvious core count comes in second to core speed (that dude with the q66 that could still do bigadv)

I doubt well see any opteron board that will provide OC abilities that the SR2 brings, the ability to mix and match CPUs (crucial for lowering initial cost) and the ability to use standard ddr3

My money is on intel this round (obviously) but I'm glad AMD is starting to land some good punches
Posted via [H] Mobile Device
Click to expand...

Our only hope is that ASUS comes out with something of the like for the AMD crowd...
 
You must log in or register to reply here.
Back
Top