GTX 780 vs. GTX 670

[TUCID]

Hi
2 questions.

I have one GTX 670 4Gb installed. How much will I gain in performance when changing to a GTX 780 3 Gb? Is the difference noticeable. I know there is a big difference in CUDA cores. On the other hand the 670 have 4 gb vs. 3 gb on the 780.

What does it require if I want my work station to perform regular tasks (using Photoshop or even using the browser) while rendering an image without delay?

I'm not a heavy user but I would like to get the most out of what I got.

//Thomas

[haknslash]

How heavy are your projects? Are you hitting 3 GB now? If not then what does it matter?

As far as performance between the two a 780 is going to give you a nice bump in performance having nearly 1000 more cores.

[Mar10]

Agree with Kenny, if you are not hitting 3GB now, its probably not an issue. Check out the benchmark thread for more details on the performance comparison. The GTX 780 has proven itself to be in line, performance wise, with the GTX 580 or just slightly faster. So if you look at the benchmark numbers the 670 runs it in 125~126 seconds and the 580 runs it in 80 seconds. Assuming the 780 would do it in 79~80 seconds thats a roughly a 37% decrease in rendering time. That is, if I did my math right. So if your typical rendering takes 30 minutes, it would come down to roughly 19 minutes. I added a couple other times below just to show the relative decrease in render time.

30> 19 minutes
15> 9:30
5> 3:09



http://forum.bunkspeed.com/showthrea...Bunkmark/page5

[TUCID]

Thanks for the answers. But sorry, I'm a little confused now. Everyone have told me that i need more CUDA cores to increase render speed. When i look at the Bunkmarks it looks like the GPU with the fastest rendertime is the one with fewest cores. So what is the truth here. Also i was recommended to buy a GPU with high memory- I don't know if I'm hitting 3 or 4 gb at the moment- how can I tell?

[Mar10]

Yes, it can be complicated. The idea that more Cuda Cores are better is still generally true within a given family of GPUs. For some reason, the Cuda Core is not a constant across families. So 1000 Cuda Cores in a GTX 5XX series card is not equal to 1000 Cuda Cores in a GTX 6XX or 7XX card. I have no idea why that is true but it is. This is why the benchmark thread is a great asset to have. You can look at a number of cards and see where it ranks.

I'm not sure how to tell how much memory you need. You can see the poly count in bunkspeed but I'm not sure how it exactly correlates to VRAM needed. Some people on this thread a little over the top with their thought that everyone needs the most vram they can afford, its bull. I personally want the fastest card for the money without over paying for a bunch of vram I don't need. Its all dependent on what type of models and rendering you do. If you do simple consumer electronic products like me, you probably don't need more than 3GB but if you are doing heavily detailed architectural stuff, then maybe you do. Check the poly count on some of your most detailed models and hopefully someone can chime in about how much vram is needed.

[david.randle]

We support about 5 Million polygons per GB of VRAM. If you launch defaulted to Accurate raytracing and never turn on Preview or Blend, then we support 10 million polygons per GB of VRAM.

[Alex_Ji]

My analysis should help you to make a decision.
http://forum.bunkspeed.com/showthrea...ull=1#post6262