Today I shall put the above-oft stated assertion to the test.
We hear manifestations of this claim in all manner of forms:
"Of course the 7970 smokes the 680 at high resolutions. It's the bandwidth!!!"
"660ti is a nice card, but it takes a dump at high resolutions. Look at that crappy 192-bit bus!"
"Of course the Titan isn't that great at 1080p. Have to crank up the resolution to take advantage of that 384-bit bus!"
So today, I've decided to put this pervasive theory to the test with my 670, in an attempt to see a) if it holds water, and b) if so, how much of the proverbial water does it hold
![thumb.gif]( "thumb.gif")
I'll be using the Valley benchmark, one that is known to be highly memory-bandwidth dependent. I'll be running it with all settings maxed, at two resolutions, the latter of which being precisely twice the size of the former, to wit: 1280x800, and 2560x1600.
At each resolution, I'll be running the test at stock memory speed (1502MHz), and again with a hefty overclock (1760, or (1760-1502)/1502 = 17.17%). We shall then assess the scaling (% fps increase/% bandwidth increase) provided by the additional bandwidth at each of the two resolutions. Granted, this is not the MOST optimal way of doing this sort of test, because increasing clock speed does not provide a 100% equal in every possible scenario equivalent to increasing the bus width, however ... a vast majority of the time ... it accomplishes the exact same thing.
Without further ADO, here are the results of said tests:
1280 x 800, 1502MHz: 75.0fps
1280 x 800, 1760MHz: 79.8fps
2560 x 1600, 1502MHz: 24.6fps
2560 x 1600, 1760MHz: 27.1fps
Running these math's through my high-tech supercomputer, I derive the following Bandwidth Scaling numbers:
@1280 resolution: (79.8-75.0)/75.0*100 = 6.4% delta fps / 17.7% delta bandwidth = 36.1% memory bandwidth scaling.
@2560 resolution: (27.1-24.6)/24.6*100 = 10.16% delta fps / 17.7% delta bandwidth = 57.4% memory bandwidth scaling.
So there you have it ... proof of a fairly positive nature that it is in fact true that as resolution increases, the relative usefulness of increased memory bandwidth also increases. Although the effect of doubling resolution does not quite result in a doubling of the bandwidth scaling, it's actually surprisingly close.
I would also add that number like 57% bandwidth scaling is incredibly high relative to most cards, and pretty damn solid evidence that the Kepler offerings came out with less memory bandwidth than they probably really should have. And if you AREN'T overclocking your memory because you've always heard it 'doesn't do much', well ... with Kepler ... it does A LOT.
I invite others to indulge in similar testing using other benchmarks or games, to see whether this is an across the board phenomenon, or what
![thinking.gif]( "thinking.gif")
We hear manifestations of this claim in all manner of forms:
"Of course the 7970 smokes the 680 at high resolutions. It's the bandwidth!!!"
"660ti is a nice card, but it takes a dump at high resolutions. Look at that crappy 192-bit bus!"
"Of course the Titan isn't that great at 1080p. Have to crank up the resolution to take advantage of that 384-bit bus!"
So today, I've decided to put this pervasive theory to the test with my 670, in an attempt to see a) if it holds water, and b) if so, how much of the proverbial water does it hold
I'll be using the Valley benchmark, one that is known to be highly memory-bandwidth dependent. I'll be running it with all settings maxed, at two resolutions, the latter of which being precisely twice the size of the former, to wit: 1280x800, and 2560x1600.
At each resolution, I'll be running the test at stock memory speed (1502MHz), and again with a hefty overclock (1760, or (1760-1502)/1502 = 17.17%). We shall then assess the scaling (% fps increase/% bandwidth increase) provided by the additional bandwidth at each of the two resolutions. Granted, this is not the MOST optimal way of doing this sort of test, because increasing clock speed does not provide a 100% equal in every possible scenario equivalent to increasing the bus width, however ... a vast majority of the time ... it accomplishes the exact same thing.
Without further ADO, here are the results of said tests:
1280 x 800, 1502MHz: 75.0fps
1280 x 800, 1760MHz: 79.8fps
2560 x 1600, 1502MHz: 24.6fps
2560 x 1600, 1760MHz: 27.1fps
Running these math's through my high-tech supercomputer, I derive the following Bandwidth Scaling numbers:
@1280 resolution: (79.8-75.0)/75.0*100 = 6.4% delta fps / 17.7% delta bandwidth = 36.1% memory bandwidth scaling.
@2560 resolution: (27.1-24.6)/24.6*100 = 10.16% delta fps / 17.7% delta bandwidth = 57.4% memory bandwidth scaling.
So there you have it ... proof of a fairly positive nature that it is in fact true that as resolution increases, the relative usefulness of increased memory bandwidth also increases. Although the effect of doubling resolution does not quite result in a doubling of the bandwidth scaling, it's actually surprisingly close.
I would also add that number like 57% bandwidth scaling is incredibly high relative to most cards, and pretty damn solid evidence that the Kepler offerings came out with less memory bandwidth than they probably really should have. And if you AREN'T overclocking your memory because you've always heard it 'doesn't do much', well ... with Kepler ... it does A LOT.
I invite others to indulge in similar testing using other benchmarks or games, to see whether this is an across the board phenomenon, or what
