GK110 a.k.a 780

[nikand]

So its coming. Probably just a month from now. Hopefully this thing is up for snuff with its 8-9 teraflops single precision and 6gb of ram.

http://www.xbitlabs.com/news/graphic...ame_Titan.html

[Alan Iglesias]

Ah...the next step cometh...

My reference point with GPU board rendering Speed-per-Dollar is based on the fermi-based GTX 580, of which I run two at the moment. It wll be interesting to see where this fits into the equation.

-Alan

[vlado]

From what I've seen, the GK110 chips are faster than the 580's; it remains to be seen what would be the exact performance/price ratio.

Best regards,
Vlado

[Alan Iglesias]

From what I've seen, the GK110 chips are faster than the 580's...

Ah...excellent to hear! Im waiting with bated breath...

-Alan

[dtolios]

I think the 780 won't be able to challenge the 580s for projects were 3GB of VRam are enough.
Performance is expected to be 80% that of a 690, and I would not expect it to be much cheaper than a 690 for at least some time - I also doubt the supply/demand game will drive prices down fast.

So, it will definitely be faster than a 580 - no question. So does the 690.
But costing nearly 2.5x, will be driving the $/performance too low.

[Alan Iglesias]

...But costing nearly 2.5x, will be driving the $/performance too low.

If it is twice the price of the 580 without being twice as fast, well that would be unfortunate. But knowing what I know it certainly wouldn't surprise me.

-Alan

[vlado]

I'm afraid we won't see any drastic improvements in GPU speeds anymore. From what it looks like, there will be incremental improvements and in a few years we might get cards that are two, maybe three times faster than what we have now. I don't think we can expect another jump in performance like what we got from the 2xx to the 4xx and 5xx cards.

Best regards,
Vlado

[dtolios]

79xx cards were promising 2x the OpenCL performance of the GTX 580 through some benchmarks - guess the benches were not compiled correctly...

The GTX 580 remains a tad faster than the 680 in computation, and a 690 is based on 2 680 GPUs (GK104), so the 690 is by definition a tad slower than 2x 580s (and more expensive with current prices).
The GK110 based card is rumored to be slower than the 690 - which makes sense since it has 2,688 CUDA cores which would be 87.5% of the GTX 690's 3,072 (also slower clocked than what we have in the 690).

So it is "safe" to assume that we will have a performance ladder looking something like: 2x 580 > 2x 680 > 1x 690 > 1x "Geforce-Titan"/780 > 1x 580 > 1x 680 etc.

List price for the 780 is rumored to be $899.
Probably won't win vs. 2x 580s, but you can do more 780s in the same space, will consume less power than 2x 580s (lil bit more than one tbh, those 580s are...) and ofc will have 6 GB buffer at your disposal. I think it is time for VRay RT GPU to have an option not to scale all the textures down, but have a priority list etc for some of better or uncompressed quality?