Originally Posted by specopsFI
That nai benchmark is interesting, but I'm still skeptical. There are too many fresh users posting screenshots of unknown origin for me to jump on the "issue confirmed" bandwagon. I need a reputable source and not just for this benchmark but actual gameplay being affected. I certainly didn't see a problem with my own Watch_Dogs testing, hitting 3750MB VRAM without excessive stuttering. Beyond that, I can't say.
I have no problem with this being discussed but I see a lack of critical thinking and confirmed evidence of an actual issue. The job of these GPUs is playing games as good as possible and that is what all the reviews and my personal experience are showing. The theory of disabled SMM's and their distribution is interesting though, as that could explain different results for different people. I do recall the rather small debate on GTX 780's (which also had three SMX's disabled) on whether or not there is a performance difference between a chip where the disabled SMX's are evenly distributed and a chip where all three disabled SMX's are within a single GPC (see http://www.anandtech.com/show/6973/nvidia-geforce-gtx-780-review
). Obviously it wasn't a big enough issue with GTX 780, but that is not to say Maxwell couldn't be different.
So still not convinced, but there could be something here. I'm more and more inclined to think we will get an answer rather sooner than later, though.
Hmm....I have quick a look on GTX 780 distribution and compare it to GTX 780 TI. There is no problem with Kepler series.
There are 6 memory controllers at the diagram. There is no shortage of bandwidth. GTX 780 has 4 raster engines while GTX 780 TI has 5 raster engines. You could group / refer them by GPC if you wish so.
Each of the raster engine has its own memory controllers. I think two memory controllers are underutilized for GTX 780 while one memory controller underutilized for GTX 780 TI. No idea why nvidia sell it as 384 bits even although the aforementioned memory controller is kinda unused in a sense.
If you take a look, nvidia could actually put total 6 raster engine / GPC with each connected to their own memory controller. Yield issue maybe?
You could say GTX 780 effective width is 256 bit while GTX 780 TI is 320 bit. But then we will have to look on how the GDDR5 being soldered / connected to the memory interface at the circuit board.
Maxwell V2 is diferrent. Instead of cutting Raster Engine unit as they had done for Kepler 780 series, Maxwell V2 has its SMM or SMX equivalent of Kepler being disabled.
EDIT: wording issue + some info
EDIT V2: Anyone with GTX 780 and GTX 780TI kind enough to run Nai Benchmark and check for bandwidth speed? Make sure both gpu memory frequency are the same, gpu frequency too! I wanna check if there is any difference in term of memory bandwidth.Edited by JohnLai - 1/23/15 at 3:10am