NVIDIA Kepler speculation thread

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-680

GTX 680 page now up at geforce.com. Many other pages about the GTX 680 and Kepler architecture still not up yet.

 

That test looks like it was done against a factory OC'd 7970 as well, the stock vs stock comparison will be even more one sided.

I am going to reserve judgement until I see Anand's review.

 

More like bandwidth + the radeon frontend/driver low res disease.

Huh? both stock 925 and 1000mhz results (interesting for clock2clock performance) in most graphs, and otherwise just the stock vs stock.. But with old 7970 launch drivers.

 

Scaling is only a bit better than GF110 and it's only 25-33% faster than GF110 at the 25x16 and lower benches published so far...

 

Why hasn´t AMD improved that after all these years?.

 

http://www.pcinlife.com/article_photo/gtx_680/results/total.png

Whole game list - as predicted the 7970 wins in AvP, Crysis Warhead and Metro. The 680 appears to have blown it away in Shogun II though, which is something of a reversal of fortunes in that game. The Batman scores look too high in favour of the 680 so I'm guessing that's favourable settings.

 

Average results GTX680@stock vs. HD7970@1000/5600:

1AA + 16AF
1680x1050: GTX 680 7,3% faster
1920x1200: GTX 680 5,1% faster
2560x1600: GTX 680 0,4% slower

4AA + 16AF
1680x1050: GTX 680 5,8% faster
1920x1200: GTX 680 3,1% faster
2560x1600: GTX 680 1,3% slower

8AA + 16AF
1680x1050: GTX 680 5,6% faster
1920x1200: GTX 680 3,1% faster
2560x1600: GTX 680 2,8% slower

 

Yeah, reading comprehension fail. D:

 

Can't wait for the GTX680@maxOC vs. HD7970@maxOC face-off promised by HardOCP.

Solid Perf/W comparison between GTX680 and HD 7870 would be very interesting, too.

 

looking against to GTX580 scores it seems that it is bw, sure radeon frontend seems suffer at low res

 

Some interesting bits on the warp scheduling in Kepler, from the white paper:

 

Wow, did that one leak as well?

 

Whitepaper: http://www.geforce.com/Active/en_US/en_US/pdf/GeForce-GTX-680-Whitepaper-FINAL.pdf

 

Its already available at geforce.com

 

GK104 acts as 1024CCs on some tests ~30% faster than GF110..

Spoiler

 

Because the superscalar design could not be fully utilized with scalar/1D instructions. Its the same like on the non-GF110/100 Fermis.

 

another theoretical benchmarks, waiting for H.fr's

http://www.ixbt.com/video3/gk104-part1.shtml

 

Strange that the big sites refrain from jumping the gun, now nvidia does it itself, kind of voiding the nda..

 

It does draw a as to why tweaktown's benchmarks and power consumption numbers are different from the other reviews. If you read the comments he does say that they didn't receive a card from them and he's not under NDA. Tweaktown's review shows higher power consumption for the GTX 680. So will the retail card's performance and power consumption actually be different then what's reviewed?

 

Looking at power consumption graph from pcinlife tests it looks like power band in which GK104 operates is narrower than Thaiti's, but pure average is very similar. I would like to see more games tested in that mix, especially games where AMD is doing well, like Metro 2033.

Ideally to get the best picture possible how efficiently each GPU renders frames I would like to see power graphs for games frame limited to 30 or 60FPS with 0xAA to 8xAA and various levels of AF. That would normalize power use per frame with driver overhead included. Hopefully some reviewers will take note and dive into this dark alley of extra test hours spent to make their review truly best!