Einstein GPU performance on 16x 8x and 4x lanes

Konz Andy
Konz Andy
Joined: 22 Aug 10
Posts: 7
Credit: 12449779
RAC: 0
Topic 200306

Hello

 

i searched a lot but didn´t find exactly what I want to know.

My Brother has an AMD FX 4320 and want to buy a new mainboard with 2 PCIe SLots.

He want to install 2 graficcards to claim more credit.

The problem is that the most mainboards for his processor only give 16x 4x lanes or 8x 8x for the second Pcie slot.

Does the 8x or 4x decrease the performance of the graficcard while computing boinc related workunits.

Does he really need a mainboard which delivers 2x PCIe X16 (they are a lot more expensive)

What I understand is, that the performance will decrease in 8x or 4x  while gaming operations but not while computing work units for Boinc, because while computing, there is not a lot of traffic on the Pcie bus.

is that right?

 

Greetings from Germany, KONZ Andy

 

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7023844931
RAC: 1804619

I have recent experience in

I have recent experience in operating a modern, but pretty cheap motherboard with two PCIe x16 slots, which drop to x8 when both are used.

The GPUs I currently operate in these two slots are a GTX 1070 and a GTX 1060.  The specific Einstein application I have been running is the CUDA55 "test application" or "beta" variant of BRP6.  The operating system is Windows 10.

For that specific set of conditions, the degradation in performance of the GTX 1070 suffered by the combined effects of downgrading the PCIe connection from x16 to x8, plus sharing the host system CPU, memory, disk, and I/O with the 1060 is so small I have not reliably measured it.

I think this pleasant condition is fairly specific to the current BRP6/CUDA55 application, as I recall just a few years ago that systems which employed multiple capable graphics cards suffered considerable performance loss from some element of the "sharing".  I don't think people ran persuasive experiments isolating which aspect of sharing was a problem for that generation of Einstein applications, though more than one person has posted confident claims of the great superiority of x16 PCIe in those days.

So, my tentative answer for your case, is that a person making a purchase decision now aimed a price/performance for Einstein BRP6 would not find it a good choice to spend a lot extra for double x16 PCIe slots.

We don't know what the future will bring for GPU computing at Einstein.  At least in part that depends on how well the scientifically interesting computation maps to the GPU architectures using available tools, and also how much of scarce development resources is devoted by the project to forcing or improving that fit.

Konz Andy
Konz Andy
Joined: 22 Aug 10
Posts: 7
Credit: 12449779
RAC: 0

Hello,   thank you for your

Hello,

 

thank you for your answer

he will install 2 gtx950 or gtx960 style cards. used they are really affortable.

 

i didn't well understand your explication,

will the 8x slot braking the computing performance of the card or not,

if so, he will spend a 50 bucks more too buy a 2x 16 lanes motherboard.

but if the lost of performance is minimal ( 5-10%) on the second card, 

 he will spend the money on more performance video cards,

Or perhaps an upgrade cpu wirh more cores.

Can deactivating one core of the cpu for graficcard serving increase the performance of the videocards?

 

 

greetings from Germany

Konz Andy

 

 

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7023844931
RAC: 1804619

My observation was that the

My observation was that the 1070 performance was not lower ( not even 1%) when it shared the PC with a 1060 and thus was reduced from 16x to 8x.

Sometimes using the web site preferences (or other methods) to reduce CPU usage by other applications can help GPU computing performance--sometimes not.  Experiment on your own system with the applications of current interest is the best guide on that point.

 

floyd
floyd
Joined: 12 Sep 11
Posts: 133
Credit: 186610495
RAC: 0

Konz Andy wrote:he will

Konz Andy wrote:
he will install 2 gtx950 or gtx960 style cards. used they are really affortable.

Only run a multi GPU setup if you want more computing power than a single GPU can give. One big GPU will be just as good as a pair of medium ones, and it's less effort. My advice is to keep your board and CPU and add a single GTX970 or better if Nvidia is your choice. You may have to replace your PSU, if so make it big enough to have reserves for a second GPU. That way you can still upgrade later.

Quote:
Or perhaps an upgrade cpu wirh more cores.

This one should be good enough for any single GPU.

Quote:
Can deactivating one core of the cpu for graficcard serving increase the performance of the videocards?

 You'll need to run less CPU tasks but how many are still possible depends on the type of GPU.

 

mmonnin
mmonnin
Joined: 29 May 16
Posts: 291
Credit: 3229563954
RAC: 1114559

floyd_7 wrote:Konz Andy

floyd_7 wrote:

Konz Andy wrote:
he will install 2 gtx950 or gtx960 style cards. used they are really affortable.

Only run a multi GPU setup if you want more computing power than a single GPU can give. One big GPU will be just as good as a pair of medium ones, and it's less effort. My advice is to keep your board and CPU and add a single GTX970 or better if Nvidia is your choice. You may have to replace your PSU, if so make it big enough to have reserves for a second GPU. That way you can still upgrade later.

Quote:
Or perhaps an upgrade cpu wirh more cores.

This one should be good enough for any single GPU.

Quote:
Can deactivating one core of the cpu for graficcard serving increase the performance of the videocards?

 You'll need to run less CPU tasks but how many are still possible depends on the type of GPU.

 

 

I would also recommend a single bigger card now. Which leaves you with an option to add a 2nd later for more production. Buy as big of a card as you can afford now. Pascal does offer some reduced power as well if you went that route.

If you do get one card, you'll be able to check the PCI-E bus utilization to make sure you're not close to half of it. I'd have to check my own card but I don't think E@H runs close to 50% even with 2 cards in 8x8.

Konz Andy
Konz Andy
Joined: 22 Aug 10
Posts: 7
Credit: 12449779
RAC: 0

Hello Thank you for your

Hello

 

Thank you for your answers.

 

The problem is, that he can get 2 cars for 70-80 € each.

a single card with the same performance as the 2 cards will cost a lot more than the 2 older cards.

So here is the idea to buy 2 old cards each witch 2000-2500Gflops SP each (75€ used) , instead of buying a single card with 5000 Gflops witch will cost 255€ used or 366€ new (GTX 980)

For this money he can get 2x GTX 680 with 3090 Gflops each at 119€ used and they deliver both 6180 Gflops.

A GTX 980 TI 6045 Gflops cost about 394€ new

 

Thats the reflection that we are coiurrently thinking about

 

 

 

 

 

slozomby
slozomby
Joined: 8 Dec 05
Posts: 15
Credit: 256213
RAC: 0

dont forget the cost of

dont forget the cost of power. 

1 card at 150w = 180€ a year in power @ .1372€/kWh 


so your 2 680s = (120+180)*2 = ~600€ to buy and run for a year

where as a single 980ti costs 400+180 = ~580€ to buy and run for a year. 

after the 1st year its far cheaper to run 1 faster card. 


Just my 2cents


archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7023844931
RAC: 1804619

slozomby wrote:dont forget

slozomby wrote:

dont forget the cost of power.

I agree, but with power cost and output per unit purchase cost in the picture I think people should strongly consider members of the Pascal generation from Nvidia.

In particular the newly released 3 GB version of the 1060 has models available in the USA for $200.  I'd expect the 3GB variant to offer well over 90% of the Einstein performance of the 6GB variant (despite the name it has somewhat fewer cores than the 6GB), which is a lot of performance at this price point, at very modest power consumption.

Any Pascal (1080, 1070, 1060) has much better power efficiency than any Maxwell2 (980, 970, 960) and far better than yet older generation Nvidia or AMD cards as for example the 680.

While I think the 1080 is priced too high to be price efficiency attractive, so far each newly released member of the Pascal family has been appreciably cheaper with a surprisingly moderate drop in Einstein performance.  So in this family at this time, purchase price efficiency favors the lower-ranking cards.

On the two versus one topic, it depends on what current prices are offered.  Just now, a dual 1060 box would cost much less than a single 1080 box, and out-perform it on Einstein.  Only if one eventually wished to push the box to the max by populating the second slot with a high-end card does the option of populating both slots with the more moderately price modern card look bad.

The "modern" part is very important once power costs are taken into consideration.

[Edited to add these power observation comments: I currently operate a two-card machine with one GTX 1070 and one GTX 1060.  Both are running with core and memory overclocks a couple of steps below the maximum I've observed to give correct Einstein GRP6 results at 3X (and thus are burning appreciably more power than they would running 1X with stock clocks).  The total system power at the wall is about 295W.  HWinfo reports that the 1060 card is directly consuming a little under 100W.  As it is producing BRP6 work at approximately a 143,000 cobblestone/day rate, that is very efficient.  This is a $250 GTX 1060 6GB card.  I think the $200 3GB cards would produce only slightly less, while costing only 80% as much, and possibly being slightly more power efficient.]

 

floyd
floyd
Joined: 12 Sep 11
Posts: 133
Credit: 186610495
RAC: 0

Isn't this about replacing a

Isn't this about replacing a single GT610? First I've seen dual GTX950 or 960 mentioned, where I still think a single 970 or 980 to be a good alternative, now we're at 980Ti vs dual 980 or 680. I suspect someone's getting carried away.

 

If I interpret the OP correctly, he prefers Nvidia cards and there used ones. So Pascal cards are probably no option and Maxwell(2) may be the best choice unless one finds something older real cheap. In the Maxwell2 range, there is a big gap between the 970 and 960 with respect to Einstein because of the significantly lower memory throughput of the latter. If you compare a pair of cards below that point to a single card above in my opinion the pair isn't worth it. If you compare cards above, i.e. you're approaching the top end, a pair of smaller (or less big) cards will probably give better output, but that doesn't necessarily mean they're more efficient.

 

There's been a similar gap in the Kepler range below the 660 if I remember correctly and we may still see it with Pascal. If that happens it will be somewhere below the 1060 apparently.

 

slozomby
slozomby
Joined: 8 Dec 05
Posts: 15
Credit: 256213
RAC: 0

the op (KONZ ANDY) has

the op (KONZ ANDY) has mentioned :

Quote:

he will install 2 gtx950 or gtx960 style cards. used they are really affortable.

So here is the idea to buy 2 old cards each witch 2000-2500Gflops SP each (75€ used) , instead of buying a single card with 5000 Gflops witch will cost 255€ used or 366€ new (GTX 980)

For this money he can get 2x GTX 680 with 3090 Gflops each at 119€ used and they deliver both 6180 Gflops.

A GTX 980 TI 6045 Gflops cost about 394€ new

I was just pointing out the Real cost of crunching extends beyond the initial card purchase. and trying to point out that its almost always better/cheaper in the long run to buy one faster card than 2 smaller cards using the examples he provided. 

 

so i think the consensus is 8x8 pci lanes isnt really hurting anything currently. and 1 big card is perferrable to 2 smaller cards.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.