NVIDIA Kepler speculation thread

I wonder how it does with my volume fluid demos, as GTX 580, was extremely poor at it. If someone could post a screenshot of ComputeMark, that would be appreciated too.

 

So it's 4x 48-wide SIMDs? Operating as 16x2-clock+32x1-clock with the 16 lanes having one subset of instructions and the 32 lanes having another subset?

 

http://www.computerbase.de/artikel/...ia-geforce-gtx-680/20/#abschnitt_gpucomputing
http://pclab.pl/art49195-12.html (bottom, click on test names to show)

 

6x 32-wide SIMD with super-scalar execution.
With dependent scalars you get a 2/3 troughput, like on GF104 and the other Gamer Fermi.

 

Yeah just looked at the Xbit review. It appears that once clocks and memory speeds are similar then the two cards are roughly similar. GTX 680 wins some while 7970 wins some.

As well at the overclocked settings, power consumption increases faster per mhz overclocked on the GTX 680 (~5.295 mhz/watt) than it does on the 7970 (~5.357 mhz/watt). Not big but it's there.

So, at the end of the day. Nvidia's and AMD's products are back to pre-G80 days once clocks are set to similar levels. IE - similar clocks = similar performance.

Regards,
SB

 

http://timothylottes.blogspot.com/2012/03/gtx-680-gk104.html

FXAA's creator's notes on GTX 680.

http://developer.download.nvidia.com/opengl/specs/GL_NV_bindless_texture.txt

Extension for Bindless Textures on Kepler GPUs.

 

more eyecandy -.-
http://developer.download.nvidia.com/opengl/tutorials/bindless_graphics.pdf

 

Did hardware.fr get it right? (2 scheduler pairs each assigned to 3 32-wide SIMDs + 16 SFUs + 16 LD/ST).

 

Yeah Xbit labs review looked pretty good for AMD. Once both cards are overclocked to max it's a virtual tie. Probably the most friendly to 7970 review yet.

In addition I read on Engadget today that 680's lead is slimmer in multi-GPU reviews than single, for some reason.

Still dont see a single 680 available on newegg. Glad Hardware Canucks could commend Nvidia on trouncing AMD on launch day availability.

So, when are Nvidia's down lineup cards due? I havent heard a peep, but those are what I'm actually interested in or might purchase, or purchase the AMD equivalent if Nvidia drives their prices down to better levels. Hate to say, pay 350 for a 7870 now, and have Nvidia come trounce them with something at 299 in a couple months.

I think somebody said "May". Ughh, so much waiting. And then you wait more if you want prices to settle...

AMd hasn't moved 7970 prices on newegg. So far they dont need to I suppose, with no 680's available.

 

The 7970 is blowing it away in 3 way and 4 way crossfire at eyefinity resolutions -

http://nl.hardware.info/reviews/264...sli-review-english-version-aliens-vs-predator



caveat - when the drivers work

 

There were a ton available yesterday...and they sold out very fast. I was showing stock until after lunch yesterday of about 8 different boards. Sorry you didn't notice, but, like you're branding of B3D as Nvidia territory, your lack of observation doesn't change facts.

Back to more important things, I was guessing AMD would win the hires war this round. I wonder if it's simply the memory per card? 2GB seems a bad choice IMHO.

 

OUCH!

 

At this point it probably has more to do with bandwidth than capacity.

 

I start to ask me if the difference is not coming from the use of FXAA in some tests.( Lab501 use too 4xMSAA without FXAA it seems ), even on BF3 without FXAA the 7970 stock is faster of the 680 in all resolution..

Or maybe the place tested is different, or the driver, i dont know. Well maybe i take one 680 at the end of the month for bench it, so will see.

For sure results in xbits lab and Lab501 are completely different of what we have seen in other reviews.

 

Just a side note commentary on why we continue to use the wrong term for throughput. Bandwidth should equate, in digital terms, to bus width, and what we call bandwidth should be bus width * bus clock...anyway, I realize it has become a colloquialism, but, as a scientist I always cringe.

 

Actually, our use of the term Bandwidth to describe bus width * bus clock is consistent with the Shannon-Hartley theorem, which relates channel capacity (digital bandwidth) to channel bandwidth (analog spectrum bandwidth).

http://en.wikipedia.org/wiki/Shannon–Hartley_theorem

So as a scientist, you shouldn't cringe.

 

That Metro 2033 result at 5760 must be due to running out of framebuffer surely.

I think that bandwidth is a major factor in the heavy losses in AvP and BF3 but the metro behaviour looks more like a hard limit on fps.

 

It's several years old.

 

Meh...I'm a Physicist / Materials Scientist and I'm fine with channel capacity or throughput, but bandwidth - at least in Physics - has a very specific meaning and it's not throughput or channel capacity. It is used to calculate those parameters, but it is one of many variables.

 

Thanks for the graph. What you say must be the case - 1+1+1 = 1.1? Perhaps then the AvP and BF3 results could be a combined effect of throughput (lol) and framebuffer.