GTX 580 is not as good as GTX480 for CUDA ?

Here in NL I see C2050 in webshops for <2keuro, so 4 high-end cards are probably equivalent in price. There is one benefit of Tesla becoming a bit more popular: prices will be coming down.

Here in NL I see C2050 in webshops for <2keuro, so 4 high-end cards are probably equivalent in price. There is one benefit of Tesla becoming a bit more popular: prices will be coming down.

Here in Australia using an exchange rate of 1 AUD = 0.7 EUR…
GTX480 ~ AUD 480 = 336 EUR (MSY - typical price)
C2050 ~ AUD 3232 = 2262 EUR (itsdirect.com.au, only web price i could find)
C2070 ~ AUD 5215 = 3650 EUR (yikes)
Assuming I could get the NL price you mentioned and that the AUD stays strong (typically more like 0.6 or 0.5 than 0.7) then that still gets 6 x GTX480,s for one C2050.

I think quad GTX480’s would rip apart a C2050 in all but one scenario. Assuming you can scale your code to multiple GPU’s (mine scales from 1 to 4 with 92% parallel efficiency) then you get 4 times the memory bandwidth (distributed of course) with 4 x GTX480 vs one C2050 and you also have twice the memory available (again distributed). You only get the same arithmetic throughput due to nvidias crippling drivers however it is still cheaper overall and users that don’t require DP can use the full power of your 4 x GTX480 powerhouse. Also, as far as I can understand, the C2050 is clocked at 575 MHz while the GTX480 runs at 700 MHz so you should get more DP performance from the quad GTX480’s despite nvidias best efforts! When you factor in the fact that the C2050 memory is clocked at 1.5GHz while GTX480 is 2GHz then the tesla looses by an even greater margin.

I suppose the only real use of the tesla cards is if you want to build the most powerful (GPU based) possible single computer, ie. 4 x C2070 in one box. Then you could work on a 24GB distributed memory problem beating the quad GTX480’s with 6 GB… Then again I don’t have $21,000 to spend so this scenario doesn’t really apply to me…

Here in Australia using an exchange rate of 1 AUD = 0.7 EUR…
GTX480 ~ AUD 480 = 336 EUR (MSY - typical price)
C2050 ~ AUD 3232 = 2262 EUR (itsdirect.com.au, only web price i could find)
C2070 ~ AUD 5215 = 3650 EUR (yikes)
Assuming I could get the NL price you mentioned and that the AUD stays strong (typically more like 0.6 or 0.5 than 0.7) then that still gets 6 x GTX480,s for one C2050.

I think quad GTX480’s would rip apart a C2050 in all but one scenario. Assuming you can scale your code to multiple GPU’s (mine scales from 1 to 4 with 92% parallel efficiency) then you get 4 times the memory bandwidth (distributed of course) with 4 x GTX480 vs one C2050 and you also have twice the memory available (again distributed). You only get the same arithmetic throughput due to nvidias crippling drivers however it is still cheaper overall and users that don’t require DP can use the full power of your 4 x GTX480 powerhouse. Also, as far as I can understand, the C2050 is clocked at 575 MHz while the GTX480 runs at 700 MHz so you should get more DP performance from the quad GTX480’s despite nvidias best efforts! When you factor in the fact that the C2050 memory is clocked at 1.5GHz while GTX480 is 2GHz then the tesla looses by an even greater margin.

I suppose the only real use of the tesla cards is if you want to build the most powerful (GPU based) possible single computer, ie. 4 x C2070 in one box. Then you could work on a 24GB distributed memory problem beating the quad GTX480’s with 6 GB… Then again I don’t have $21,000 to spend so this scenario doesn’t really apply to me…