Printable View
In the continuing curiosity killed the cat series... Once upon a time I used to crunch with some cuda crunchers, which put out enough heat to melt the arctic, and sucked dam generators like Bonneville dry. So what is the state of the cuda cards now? What is the best card they put out now when viewed from the least heat and power used aspect yet is a respectable cruncher when viewed from a nvidia perspective and not ranked against amd/ati??
The best CUDA card with the least heat, yet a respectable cruncher might be the GeForce GTX 950.
With a tdp of 'only' 90 Watt it gives 1572 GFLOPS Single Precision and 49.1 GFLOPS Double Precision, or 1.5x the SP of a HD 4850 and 0.25 the DP of that card -for a slightly lower tdp.
As a bonus you get a much more recent design.
If you can spare $5.000, buy a Tesla K80.
wow.............. :-(
It's not all that bad, take a GeForce GTX 750 Ti:Quote:
Originally Posted by Brucifer
1306 GFLOPS SP, 40.8 GFLOPS DP and 60 Watt tdp
Explain the 'Titan Z' Dirk... :blob3:
http://www.nvidia.com/gtx-700-graphi.../titan-z-2.jpg
The Titan Z is the utter king of consumer CUDA -like the Tesla K80 is for the professional.
It has compute capability 3.5, which is better than the -later- compute capability 5.0 and 5.2 of e.g. the Titan X.
It has 8122 GFLOPS SP and 2707 GFLOPS DP, but burns your wallet down with a tdp of 375 Watt.
Titan X does 6144 GFLOPS SP and 192 (! no more) GFLOPS DP, burning at 250 Watt tdp
At 375w 24x7 that's going to put a noticeable dent in one's monthly power bill, not to mention the amount of heat that thing has to throw off. Where would that card excel at in the DC crunching world? And at what point in the nvidia DP game does it make more sense to go to the professional cards?
Definitely interesting from the power consumption view. So we have nvidia and amd/ati putting out cards. As I understand it from a programing viewpoint, the nvidia cuda and amd programming approaches are different...????? I guess what I'm wondering is what are the strength/weaknesses of each and why would a project select one versus the other ie, what are the considerations involved? Any idea?Quote:
Originally Posted by Dirk Broer
If you were going to tackle learning to program gpu's, which would you tackle and why?
And another related question, regarding both brands, do the mfr's utilize that same methods for all their cards, or are there variations within each mfr's card designs which utilize different toolkits, etc.??
It would excel in those (sub)projects that have DP as prerequisite (MilkyWay@Home as an example). For us consumers there is no need to go for the professional (Quadro, Tesla) cards unless we have money to burn.Quote:
Originally Posted by Brucifer
It is what the programmer of the project wants/has learned that decides which API is being developed in.Quote:
Originally Posted by Brucifer
I would try to make the OpenCL API better than both CUDA and CAL++/Brook, perhaps by integrating the API with OpenGL too -if possible.Quote:
Originally Posted by Brucifer
nVidia used to have only CUDA, but the last generations also support OpenCL. Ati used CAL++/Brook and shortly after -or before- AMD acquired them also started to support OpenCL.Quote:
Originally Posted by Brucifer
nVidia still continues CUDA, AMD has stopped CAL++/Brook.
My first real GPU cruncher was a 2nd hand nVidia GeForce GTX 260, which has come down with a broken fan a year or so ago, much to the chagrin of the power company.
It is a lot cooler in the man cave too, nowadays. How good/bad are these nVidia cards over the years? I will compare the GTX x60 series for you.
* -a, -b and -c are my designations, there are not to be found on the box of the card. You'd best check with GPU-Z what actual model you have of such a card.
Model/Type GPU Fab in nm Shaders TMU ROP GFLOP SP GFLOP DP TDP Gflop(SP)
/WattGflop(DP)
/WattTesla 1/8 of SP Fermi 1/8 of SP Kepler 1/24 of SP Maxwell 1/32 of SP Pascal 1/32 of SP
Note that the OEM cards are best left alone (with the exceeption of the GTX 660). Do not buy a system with such a card included, make your own systems containing the 'real' cards, they offer better value.
Note also that the best DP performer, the GTX 560 Ti-c, does so at terrible efficiency. If you have the money to buy a card in this class and you want to go CUDA, buy a GTX 1060-b with GP106-400 GPU. Amazing (compared with the GTX 260) SP performance and rather efficient DP performance, be it at 1/32th (!) of the SP performance.
Wow, great table Dirk... interesting to see the very last and newest one is the first one to reach the 1Gf/Watt ratio
There are two ways of going broke with GPU computing. One is by buying a card that you can't afford to pay, the other is by buying a card which power needs are such that you can't afford to pay the power-bill.
Just as with people that write about cars, who have the tendency to describe the latest Ferrari's, Lamborghini's and other unobtainable sports cars, I will now describe which nVidia-cards to avoid like the plague, however good performers they are. They will let you go broke both ways!
*=prosumers only
Model/Type GPU Fab in nm Shaders TMU ROP GFLOP SP GFLOP DP TDP Gflop(SP)
/WattGflop(DP)
/WattTesla 1/8 of SP Fermi 1/8 of SP* Kepler 1/24 of SP* Maxwell 1/24 of SP* Pascal 1/32 of SP*
The first at an affordable power usage -for nVidia- that is...Quote:
Originally Posted by NeoGen
Nice Posts Dirk, thanks for sharing