Has nVidia really crippled the GTX680 CUDA performance?

[Ken1171]

A lot of people are considering purchasing the GTX680 because of CUDA performance, so I thought this could be something important to clear out. By the time I bought my GTX580, online reviews were showing CUDA benchmarks where it could outperform the new GTX680 because of "issues with nVidia drivers". Some would even imply it appears that nVidia had intentionally crippled the GTX680 CUDA performance so that it won't compete with their Quadro line of professional cards, considering it packs 3X the number of streaming processors but still gets outperformed (in CUDA) by the older GTX580. nVidia claims this would be fixed with driver updates, but it has been a long time by now.

Can someone here [who owns a GTX680] confirm or deny that?

[t_3]

the bottom line: a gtx 580 is still a good amount faster than any single gpu 6xx card if it comes to rendering. if you call it crippled depends on your pov. 6xx cards are great for games, use less power, but the design is not optimal for most gpgpu applications. if you really want power and are willing to pay the price the gtx titan (based on the gk110 chip design) is the way to go...

[t_3]

ps: afaik the quadro line based on the gk104 chip is also not faster - at least in octane. imo the critical point for nvidia is double precision performance, where they want to differ between pro and consumer; could be that there are differences between quadros and consumer cards regarding this, but since octane only needs good sp performance, this makes no difference for rendering...

[acc24ex]

..the good thing is that 5xx series is still worth their while (so I didn't have to spend more money) - the bad thing is that there was no upgrade for the past year, really now nvidia - are you playing a dirty game there and trying to force GPGPU users to buy their pro spec graphic cards - at pro prices? I mean, I did buy into 4 most expensive cards that came out 480, 580 590 gtx.. and now I will buy the new one only when it proves at least 4x580 speed

[Jaberwocky]

rumor around the net has it that the GTX780 , due to be released in June is going to be based on a cut down Titan Card.. ..just have to wait and see if this is true.

[Ken1171]

the bottom line: a gtx 580 is still a good amount faster than any single gpu 6xx card if it comes to rendering. if you call it crippled depends on your pov.

Considering I bought the GTX580 last year mainly for rendering, I now think I made the right decision, and it really shines in CUDA performance, even if only packing 512 streaming processors - a THIRD of what ships with the GTX680. It was a difficult decision because both 580 and 680 were being sold by the SAME price. The 680 was actually cheaper because there wasn't a model with more than 1GB at the time, while my 580 "classified" shipped with 3GB, plus an oversized cooling fan that keeps it from cooking itself during renders.

Besides having to plug 3 power cables and the respective powerhogging, I recommend this older GTX580 Classified for those who want to use this Octane plugin and all the other GPU renderers out there. It already comes overclocked from factory, and max temperatures are lower than 66C during full load CUDA rendering. I didn't buy this for games, but I could still play FarCry 3 and Crysis 3 at max settings without a stutter.

I am sure the GTX680 could do better in games, but it appears that it won't beat the 580 in GPU rendering performance, based on the feedback collected here. Just be aware that the 580 suffers from rather severe CUDA texture map slots limitations, where your scenes cannot have more than only 32 grayscale maps - no matter how much VRAM memory you may have. The 680 offers twice that number of slots, but you pay with lower CUDA performance.

[linvanchene]

edited and removed by user

[t_3]

For example when using adobe after effects would double precision be needed?

Can anyone point me to a forum post / blog entry etc that discusses in detail the different types of Nvidia cards (Geforce, Quatro, Tessla) and what applications they are used for?

- - -

i have no post at hand, but i can give you a short rundown:
use for pro cards like quadro/tesla: they have usually advanced drivers for specific host; say a specific max or maya driver which speeds up the internal viewport.

floating point performance: lots of other gpgpu uses, in particular scientific applications need double precision math, so this is of course an important measure in this area. gpu based raytracing like done in octane do only need single precison math.

other differences: the pro cards run usually on lower speed, thus consuming less power, generating less heat and should allow to pack more of them into a given space. and of course they give you mor vram.

so if your main interest is octane, and your host of choice is not sped up by special nvidia drivers, you won't care about quadros or teslas - at least if you don't plan to cram seven of them into your workstation (what would need single slot cards)