Remember EVGA's SR2 motherboard?

[John P. Myers]

Here's a few details on the new SR3 (http://forums.legitreviews.com/about37468.html)

[DAD]

Nice mobo, but yet again another mobo with the pic slots too close. For 4 gpus, you coud do every other slot, but 7 tesla cards would fry from heat lol

Mobo makers need to assume all pci cards are double width, and need 1-2 more slots in between for proper cooling. this would make for longer mobos and larger cases though

[DrPop]

Man, that is insane! You would definitely have to use extension cables to get 7 GPUs hooked up. I never got this straight though - does the OS have a limit of 4, or is that just a crossfire/SLI limit?

[Duke of Buckingham]

Yeha I know. I will put them in boilling wather to have some spagheti

http://allcommsolutions.com/images/cable_confusion.jpg

[DAD]

You can put 7 or 8 nvidia gpus of any type in a system, but sli is limited to 4 max. I knw you can chain 4 dual gpu cards, but don't think the driver supports beyond 2x2 cards for a max of quad sli. But for boinc purposes, you can use as many gpus as the mobo supports (I think boinc has a 32 CPU and 32 gpu limit if Im correct)

A 64bit os would be required so all that video memory could be properly mapped and addressed :) the cooling system would HAVE to be liquid IMO. And probably a2-3k watt psu, but teslas don't burn power like gaming gpus. I think a tesla takes about 150-200Ws. Keep in mind most electric heaters are 1500-2000 watts. So a system like that would heat up a room. You electricity costs would jump. But not so much for running the computer but running the computer and AC needed to cool the room lol

Nvidia does have limitations. I know gf and tesla don't mix, but multi tesla card systems often use 1 quadro for display and teslas for science crunching.

[Mike029]

You can put 7 or 8 nvidia gpus of any type in a system, but sli is limited to 4 max. I knw you can chain 4 dual gpu cards, but don't think the driver supports beyond 2x2 cards for a max of quad sli. But for boinc purposes, you can use as many gpus as the mobo supports (I think boinc has a 32 CPU and 32 gpu limit if Im correct)

A 64bit os would be required so all that video memory could be properly mapped and addressed :) the cooling system would HAVE to be liquid IMO. And probably a2-3k watt psu, but teslas don't burn power like gaming gpus. I think a tesla takes about 150-200Ws. Keep in mind most electric heaters are 1500-2000 watts. So a system like that would heat up a room. You electricity costs would jump. But not so much for running the computer but running the computer and AC needed to cool the room lol

Nvidia does have limitations. I know gf and tesla don't mix, but multi tesla card systems often use 1 quadro for display and teslas for science crunching.

Hey Dad, I think this could be the setup your looking for. =)) Dual processors, four 580's or 2 590's. This machine would putout over 2 to 3 million credits a day.

Build 2 of these and your cooking with grease.
If I remember correctly Zombie or Max build these things weekly.

Tesla GPU Workstations cost a bundle.

[DAD]

Yeah I'd have to hand build a tesla worksttion. There's no need to buy a 4K telsa card... all the C2xxx line of cards are the same gpu, same speed, etc just with more memory, and a $1999 "low end" c2xxx card is SO worth the $$ as it will output 500-1000Gflops vs only a few Gflops for a 580/590.

[John P. Myers]

the C2050/2070 put out 1030.4 GFLOPS SPFP and 515.2 GFLOPS DPFP
a GTX 580 puts out 1581.1 GFLOPS SPFP and 197.6375 GFLOPS DPFP

considering the price and performance difference, the 580 is by far the obvious choice.

Edit: I had 2 GTX 590s in my system. Was getting around 2.25 to 2.5 mil credits/day with them (had a GTX 460 too). Of course that was before PG completely nerfed everything. 4 580's are quite a bit faster than 2 590's though. :)

[DAD]

this is what I was looking for:




GTX 460
GTX 470
GTX 580
Tesla c2050


Cores
336
448
512
448


Core Clock Rate
720 MHz
607 MHz
772 MHz
575 MHz


Memory Size
1 GB
1.28 GB
1.536 GB
3 GB


DP Floating Point Speed
.25 flop/cycle
.25 flop/cycle
.25 flop/cycle
1 flop/cycle


Sample Card
EVGA 01G-P3-1370-TR
EVGA 012-P3-1470-AR
EVGA 015-P3-1580-AR
NVIDIA Tesla C2050


Cost
$180
$239
$510
$2000



Figure of Merit
(cores*clock/$)
1.344
1.138
.775
.103


DP Figure of Merit (DP*cores*clock/$)
.336
.285
.194
.103

[John P. Myers]

this is what I was looking for:




GTX 460
GTX 470
GTX 580
Tesla c2050


Cores
336
448
512
448


Core Clock Rate
720 MHz
607 MHz
772 MHz
575 MHz


Memory Size
1 GB
1.28 GB
1.536 GB
3 GB


DP Floating Point Speed
.25 flop/cycle
.25 flop/cycle
.25 flop/cycle
1 flop/cycle


Sample Card
EVGA 01G-P3-1370-TR
EVGA 012-P3-1470-AR
EVGA 015-P3-1580-AR
NVIDIA Tesla C2050


Cost
$180
$239
$510
$2000



Figure of Merit
(cores*clock/$)
1.344
1.138
.775
.103


DP Figure of Merit (DP*cores*clock/$)
.336
.285
.194
.103



So according to this table, you can pretty much buy 4 GTX 580s for the price of 1 C2050, giving you 790.55 GFLOPS DP, vs only 515.2 GFLOPS :)

The reason the pricing on the Tesla cards is so incredibly disproportional is because you're paying for unlimited tech support, which ppl like us would never need, so it's all wasted money.

[Fire$torm]

Question: How does power consumption for one Tesla c2050 compare to four GTX 580s?

Answer: According to nVidia the c2050 is rated at < = 238W. A single 580 has a max. power draw of 244W. Both cards use one 6-pin & one 8-pin PCIe power connectors which allows for a max. power draw of 300W (75W [PCIe Bus] + 75W [6-pin connector] + 150W [8-pin connector]).

Bottom Line: x4 580 = 976W max. So depending on utility rates in DAD's area the c2050 could pay for itself in 16~24 months. That's just a guess on my part.

Sorry JPM. I had to stick my nose in... :D

[DrPop]

Holy smokes! That is a lot of juice. At rates like I'm paying (SoCal business rates), that would pay for itself even faster than that. X_X

I think it comes down to how long will you run them, vs warranty (for example, I've had 2 decently high end cards crap out in less than a year now - granted, that was crunching them nearly 24/7 and not always low temps either) :D... vs how many rigs do you want? It would be VERY hard to cool 4 GTX 580s in one rig. Sure, could you do it? Yeah...would you want to? No...
My guess is it would take you at least 2 rigs to do that right. So that is extra cost...

But then...if you did get 4 GTX 580s, then you would have 2 high end CPUs to crunch with, so...???:confused: I dunno. I guess it's all a moot point for me because it's so far out of my price range right now, LOL! :cool:

I like watching your "academic" discussion, though - because you never know at what point the practice will just take off like a rocket here, and then it WON'T be out of my range. :D:o

[Mike029]

Question: How does power consumption for one Tesla c2050 compare to four GTX 580s?

Answer: According to nVidia the c2050 is rated at < = 238W. A single 580 has a max. power draw of 244W. Both cards use one 6-pin & one 8-pin PCIe power connectors which allows for a max. power draw of 300W (75W [PCIe Bus] + 75W [6-pin connector] + 150W [8-pin connector]).

Bottom Line: x4 580 = 976W max. So depending on utility rates in DAD's area the c2050 could pay for itself in 16~24 months. That's just a guess on my part.

Sorry JPM. I had to stick my nose in... :D

Here in NY that would make the difference.

[John P. Myers]

Question: How does power consumption for one Tesla c2050 compare to four GTX 580s?

Answer: According to nVidia the c2050 is rated at < = 238W. A single 580 has a max. power draw of 244W. Both cards use one 6-pin & one 8-pin PCIe power connectors which allows for a max. power draw of 300W (75W [PCIe Bus] + 75W [6-pin connector] + 150W [8-pin connector]).

Bottom Line: x4 580 = 976W max. So depending on utility rates in DAD's area the c2050 could pay for itself in 16~24 months. That's just a guess on my part.

Sorry JPM. I had to stick my nose in... :D

haha no worries. Fair is fair. While we're speaking of fair though, that was only a comparison of DP speeds. SP comparison would be 6324.4 GFLOPS vs. 1030.4 GFLOPS. Over 6x increase for only 4x the hardware :)

[DrPop]

Oh wow. So, do we have a list of which projects are single and which are dual precision? Seems like the question of "best" rig for crunching is actually quite complex. :p