Kepler vs Fermi performance

[deifgiri]

Hi mates,

I know it's a common issue here, but it's about GTX cards again, I just bought 2.0 version, I'm using a hacktintosh (OSX 10.9 Mavericks and C4D plugin) with 2 low end GTX cards right now, GTX580 and GTX 60, I'm quite happy with overall performance and render speeds.
I'd like to setup another machine, this time with higher CUDA cores cards, I'm going to GTX680 (1500 cores) and GTX780 (2300 cores) but I've been searching and reading on the internet about benchmarks and tests, and I still haven't figured out which performs better if new kepler technology or old fermi? I saw benchmarks (Tom's Hardware) of 580 (500 cores outperforming 680 (1500 cores) so my question is simple, in the last 2.0 version kepler is optimized and scales exactly as more cuda cores the cards have?
So in theory GTX680 has to be 3 times faster than GTX580, right?

thanks,

Dave

[mib2berlin]

So in theory GTX680 has to be 3 times faster than GTX580, right?

Oh no, the GTX 580 is still faster.
Here is my benchmark with another Cuda render engine.
It is not Octane but should be very similar:

http://blenderartists.org/forum/attachm ... 1400534233

Iirc, it is not an actual spreadsheet about performance of Version 2.0 with different cards available.

Cheers, mib

[gabrielefx]

My GTX680 is faster than my old GTX580 specially using pmc
remember that the GTX680 can handle 144 rgb textures, the GTX580 only 64.

regards

[mib2berlin]

gabrielefx, could you say how much 680 is faster, not 3 times?

Cheers, mib

[gabrielefx]

slight faster...

Titans first edition are 1.8-1.9X faster (with overclock) than GTX680s

[deifgiri]

Octane's official FAQ says this:

What is the best/recommended way to increase Octane Render's performance?

As Octane Render completely relies on the GPU for rendering and does not use the CPU (besides scene loading), a more powerful graphics card (or multiple GPUs, see next question) will be required to increase Octane's rendering performance, not adding more CPU cores. Octane scales linearly with the number of CUDA cores within a given GPU architecture (e.g. the GTX 690 with 3072 CUDA cores is twice as fast in Octane as the GTX 680 with 1536 CUDA cores.


So the more CUDA cores the the faster, but some Benchmarks I've seen don't show exactly this. (GTX580 faster than GTX680 for exemple)

I'm now a bit confused about this 'scale linearly' concept.

[prehabitat]

Octane's official FAQ says this:

..... Octane scales linearly with the number of CUDA cores within a given GPU architecture (e.g. the GTX 690 with 3072 CUDA cores is twice as fast in Octane as the GTX 680 with 1536 CUDA cores....


So the more CUDA cores the the faster, but some Benchmarks I've seen don't show exactly this. (GTX580 faster than GTX680 for exemple)

I'm now a bit confused about this 'scale linearly' concept.

so within the GTX 600/700 series cards(Kepler Architecture); CUDA cores are 'equal' (for these purposes*) and therefore 3000 is twice as fast as 1500 - it scales linearly...
Within the GTX 500 series(Fermi Architecture) 1500 is 3 times faster than 500 cores.

*in truth, 3000 cores running at 800mhz are not exactly twice as fast as 1500 at 1200mhz. but as far as I know you could multiply the # of cores, by the Ghz and get a better estimate. so using the example above:
3000 x 0.8Ghz = 2400 Cuda Power Cycle Units (I made this name up)
1500 x 1.2Ghz = 1800 Cuda Power Cycle Units.

Things like the above is why you see differences WITHIN the same architecture (within Kepler for example).
BETWEEN different architectures(Fermi vs. Kepler), there are far more things at play; which require a better understanding of hardware. at any rate, I recommend a gtx770 4gb as a cheaper; Entry-level card. and gtx780ti as a good top-end card.

[mark0spasic]

Look at this graph: