GK110 = SLI 680s, performance projections based on Tesla specs

[Ranguvar]

Team OCAHOLIC benchmarked a GTX 680, 670, and double 680s, and then made performance projections of GK110 based on specs from the Tesla G20 whitepaper.

The results? Slightly better than SLI 680s.

Now, of course, these are projections. But this chip probably will become the GTX 780 next year, and it's interesting to think on this now.

Thoughts?

English source: http://videocardz.com/33208/nvidia-kepler-gk110-performance-preview
Original source: http://www.ocaholic.ch/xoops/html/modules/smartsection/item.php?itemid=747

The ”if rule”: So if the drivers perform like the 301.25 we used. If we didn’t do any mistakes in the calculation and the prediction. And if the scaling works well like we analyzed with the GK104 then this card is a monster in performance. We don’t know the price, power consumption and the heat dissipation but what we can tell, if our results are very accurate, is that this GK110 is powerful. It will be around 1.8 times faster than a GTX 680 in synthetic benchmarks and 1.6 faster in games. In fact it can be compared directly with the GTX 680 SLI as in average it performs around 3-5 % better.

We remind you that those aren’t the real results of the GK110 performance nor random numbers, it is a prediction based on a valable hypothesis!

 

[Jorona]

Good ol'speculation. I speculate it will come with green lettering one the cooler and at full load, it'll be loud.

 

[xoleras]

I can't wait until Q1 2013 to see? I seriously don't see whats the point to this, the graph on that website is complete speculation......not real performance data

 

[jojo69]

Good ol'speculation. I speculate it will come with green lettering one the cooler and at full load, it'll be loud.

rofl

at least they did base their speculation on some real analysis...but yeah

 

[xoleras]

rofl

at least they did base their speculation on some real analysis...but yeah

Taking core count and multiplying 3dmark performance by 'x' doesn't equal analysis. Performance doesn't scale in such a fashion. Your 3dmark11 score doesn't double when you SLI despite having 2 times the cores.....so this entire thing is pretty ridiculous to be honest.

 

[IdiotInCharge]

Taking core count and multiplying 3dmark performance by 'x' doesn't equal analysis. Performance doesn't scale in such a fashion. Your 3dmark11 score doesn't double when you SLI despite having 2 times the cores.....so this entire thing is pretty ridiculous to be honest.

I don't think that you can compare SLi to doubling the core count. If memory bandwidth scales with the core count, it could very well hit 2xGTX680 levels.

 

[xoleras]

I don't think that you can compare SLi to doubling the core count. If memory bandwidth scales with the core count, it could very well hit 2xGTX680 levels.

Not exactly. Keep in mind that with larger die sizes come compromises, mainly in terms of heat / TDP which will adversely affect clock speeds. I'm sure the chip will perform great, but very not likely to be double.

 

[IdiotInCharge]

Not exactly. Keep in mind that with larger die sizes come compromises, mainly in terms of heat / TDP which will adversely affect clock speeds. I'm sure the chip will perform great, but very not likely to be double.

Definitely right about the clockspeeds, but:

The GK104 is mid-range silicon, nowhere near the size of a GTX580's GF110. Nvidia knows how to make a big GPU, and an efficient GPU. Their work with the GTX690 is further proof.

 

[Trimlock]

With this core it is also speculated that it will more computer logic on the core as well which also affects gaming performance (not a WHOLE lot though, but does increase TDP).

The core count on the chip isn't double either, from the sounds of it the GPU itself will have immense performance and destroy competition at compute operations as well.

also remember that because if it had double the shader cores, with the exact same logic and same speed the scheduler would have to be damn near perfect to get exactly double the performance. This is one of the reasons why the 2900XT failed so hard, the scheduler sucked with that many shader cores.

 

[Ranguvar]

also remember that because if it had double the shader cores, with the exact same logic and same speed the scheduler would have to be damn near perfect to get exactly double the performance. This is one of the reasons why the 2900XT failed so hard, the scheduler sucked with that many shader cores.

To me this sounds like a perfect case for Kepler then. Kepler has a super-simple scheduler and is designed around this, so that sounds to me like that logic was put elsewhere.

Or maybe I'm wrong. Good chance there too.

 

[Trimlock]

To me this sounds like a perfect case for Kepler then. Kepler has a super-simple scheduler and is designed around this, so that sounds to me like that logic was put elsewhere.

Or maybe I'm wrong. Good chance there too.

How does a scheduler become simple when its applying instruction out to 2880 shader cores? I was under the impression that the scheduler in kepler is pretty impressive in complexity.

 

[Xeth]

1. GK110 probably won't reach the same clock speeds as GK104.
2. If the yields for full functioning GK104's are this bad, there's no way they're going to have enough "perfect" GK110's, and if they do, those will be reserved for the HPC market where they can be sold at $3-4k a piece. More likely the gaming version will come with 1 or even 2 SMX's disabled (2688 or 2496 cuda cores).

 

[sirmonkey1985]

lol using a tesla card specs as a comparison to a gaming card.. wow, desperate for people to view your site eh?

there is a reason we don't use tesla cards for gaming, they aren't designed for it and thus trying to use their specs as a comparison to a gaming card is just stupid.

i'm also leaning toward your guesses as well Xeth.

Definitely right about the clockspeeds, but:

The GK104 is mid-range silicon, nowhere near the size of a GTX580's GF110. Nvidia knows how to make a big GPU, and an efficient GPU. Their work with the GTX690 is further proof.

what is a GTX 690 proof of? that they can slap 2 gpu's on a card and call it a GTX 690? sorry to break it to you but they did that with the GTX 590, and the GTX 295(single pcb version).. nothing new and thus proof of nothing except that you fell for the marketing numbers on it. its just 2 GTX 680 GPU's on one card, nothing special.

 

[hyperion0101]

The biggest benefit IMO of the 2880 shader beast IMO isn't the performance compared to a 690. It's that there won't be the multi-GPU issues with SLI, which still exist. You'll get performance better on a single GPU.