NVIDIA Kepler speculation thread

The take-home-message is that Cypress basically achieved an average of +70% real-world gaming performance @ only 20% more bandwidth when compared to RV770.

Now imagine where GK104 could possibly end up given a similar evolution - when GF114 used 2Ghz GDDR5 ...
 
GTX285 -- 166 GB/s
HD5870 -- 153 GB/s

;)
You are right, I took 280's value by mistake. Anyway, the bandwidth of GTX 285 and HD 5870 is almost identical, while HD 7970's should be about 50% higher compared to a GPU with 256bit MC. It's really hard to imagine that such a product could be significantly faster than HD 7970. No process advantage, no bandwidth advantage, yet it should be smaller, faster and less power-demanding? It would be great, but it sounds overly optimistic to me.
 
Remember Matrox Parhelia with it's 20 GB/s of bandwidth - and how it compared to a Radeon 8500 or GF4 Ti 4200?
 
Remember Matrox Parhelia with it's 20 GB/s of bandwidth - and how it compared to a Radeon 8500 or GF4 Ti 4200?
Parhelia didn't had any compression for the frame-buffer data or depth occlusion culling -- the wider bus in their solution was a borderline "sham". :???:
 
Parhelia didn't had any compression for the frame-buffer data or depth occlusion culling -- the wider bus in their solution was a borderline "sham". :???:

I know - all I wanted to say is: memory bandwidth alone does not make a chip fast or slow. It's how much bandwidth the architecture really needs to saturate.
 
I know - all I wanted to say is: memory bandwidth alone does not make a chip fast or slow. It's how much bandwidth the architecture really needs to saturate.

Judging by the scaling I've seen from memory overclocking, I'd say Tahiti is fairly well balanced. The amount of memory bandwidth on R600 was ridiculously high, as evidenced by RV670.
 
My guess is that there is no GK104 (like Neliz hinted) and only GK110. Maybe it was just a way to catch leakers?
When was the last time NV had only 3 GPUs in a line up?
Is it possible that due to the current state of 28HP output and AMD's Tahiti price positioning NV simply doesn't have any pressure to launch Keplers soon?
 
What could be impressive, if such feat by Nvidia was really achieved, would be the performance per transistor. Tahiti has 50% more with 50% more transistors (more or less) so the ratio is the same of the previous generation, which is not bad if we consider there are new features and improved general computing capability. But if the GK1XX is indeed smaller and faster than Tahiti (which has about the same transistor count as GF110) then we would have quite an improvement in this metric by Nvidia. Possibly this could be reached by architecture overhaul (but it would be interesting to know how) and maybe by having very high clocks while keeping the thermal envelope low. We'll see, in the recent history from Nvidia normally there was an increase in power usage with each new architecture (GF110 is a "refining" of GF100 so I'd not count it as a "new architecture") but who knows...
 
When was the last time NV had only 3 GPUs in a line up?
Is it possible that due to the current state of 28HP output and AMD's Tahiti price positioning NV simply doesn't have any pressure to launch Keplers soon?

GT200 and GT200b say hello?
There wasnt a GT204 or GT214 right?
And the GT216 and GT218 only showed up very late.
 
GT200 and GT200b say hello?
There wasnt a GT204 or GT214 right?
And the GT216 and GT218 only showed up very late.
There was G98/B, G94/B and so on. Considering that all of them including GT200 were on the same process you could argue that GT200 was a part of G9x line up. By the way, it may well turn out to be pretty similar this time between GK110 and GK10x parts...
 
There was G98/B, G94/B and so on. Considering that all of them including GT200 were on the same process you could argue that GT200 was a part of G9x line up. By the way, it may well turn out to be pretty similar this time between GK110 and GK10x parts...

True, but the point still stands: there wasnt ANY GT2x4 part, just as there might not be a GK104 either ;)
Looking at S.I., you can see that the upgrades to the lower tiers 78xx and 77xx (e.g. rumours putting 7770 around 6850 level) are probably not going to be that much faster than its predecessors, so a GF114 might still be enough to compete for a while, since we will probably have 2 families on 28nm, namely "Kepler" and "Kepler refresh". The chip for the GK1x4 tier might come later then.
 
True, but the point still stands: there wasnt ANY GT2x4 part, just as there might not be a GK104 either ;)
Yeah. And there might not be a Pitcairn. It's all a dream, Neo.

Looking at S.I., you can see that the upgrades to the lower tiers 78xx and 77xx (e.g. rumours putting 7770 around 6850 level) are probably not going to be that much faster than its predecessors, so a GF114 might still be enough to compete for a while, since we will probably have 2 families on 28nm, namely "Kepler" and "Kepler refresh". The chip for the GK1x4 tier might come later then.
I'm not so sure that Kepler's "mid-gen kicker" (or whatever Huang called that) will come on 28nm. 20nm production is supposed to start at the end of this year. Considering that Kepler will launch sometime between now and the end of 2nd quarter it's quite possible that next line up will consist of Kepler-based GPUs on 20nm and will launch in a second half of 2013. GF11x line was NV's fixing of GF10x line on 40G. If GK1xx are good right from the start then there is no reason for them to fix them on the same process tech.
 
GK110 A2 in the wild:

http://diybbs.zol.com.cn/56/231_558125_49.html

A2 revision is in line with what Charlie said.

My guess is that there is no GK104 (like Neliz hinted) and only GK110. Maybe it was just a way to catch leakers?

looks like another fake..

GK and bottom A2's 2 is bold and upper A2's 2 is bigger.. also numbers on the fake picture looks very similar to the original one..

http://www.forum-3dcenter.org/vbulletin/showpost.php?p=9131536&postcount=3178
 
VR-Zone update: http://vr-zone.com/articles/report-nvidia-gk104-branding-specifications-revealed/14589.html

GK104 impressively doubles up on GF114's core count to 768 CUDA cores - although Kepler shaders will be different from Fermi counterparts. Single precision performance is rated at above 2 Teraflops, twice that of GTX 560 Ti and over 50% higher than GTX 580. GK104 will continue to feature a 256-bit memory interface, but with frame buffer doubled to 2GB presumably at higher clocks. With everything increasing, unfortunately so does power consumption, to a TDP of 225W. A GTX 660 Ti variant is also expected in the future.
————————

The comparisons to GTX 560 Ti and GTX 580 seem off: 2x GTX 560 Ti = 2.53 TFLOPS and 1.5x GTX 580 = 2.37 TFLOPS. I understand these numbers are technically "above 2 Teraflops," but if the GTX 660 is indeed around those numbers, why not just say "around 2.5 teraflops"? Maybe they're hinting at GTX 660 being close to 2.5 TFLOPS? Or they messed up on the numbers somewhere.

As for clock speeds, if the Kepler CCs are 2 (SP) FLOPS/cycle like Fermi's (it doesn't clarify how "different" the CCs are), then that would give over 1.3 GHz to get over 2 TFLOPS. That is, unless the GTX 660 has disabled CCs like the GTX 560.
 
Sounds like they are saying 2x the 560 rather than 2x 560 Ti. Which makes the performance even more impressive - assuming they have the capability to scale up to a 660Ti and far beyond.
 
Back
Top