G92b GTS 250 memory bandwidth starved?

[Deanodarlo]

I have a E-Green GTS 250 in my secondary HTPC, with 128 shaders and clocks at 675 1456 and 256bit GDDR3 mem at 900.

I use it for playing the occasional game on the LCD TV downstairs and with the TV's scaling, even low resolutions actually look pretty good compared to a PC monitor on my gaming rig. I was surprised at the pic quality and even resolutions as low as 1024x768 look nearly as good as the PS3 at 720p due to the way the TV upscales the image.

Now I've never owned a G92 based card before and it seems to be severly limited by memory bandwidth.

I've increase the mem from 900 to 1000 (10%) and I almost get a 10% increase in performance in Lost Planet 2, OCCT and 3dmark 03/06 benchmarks!

However, if I increase the core and shader clock, I get miniscule performance gains, like less than 1 FPS. Also if I overclock too far, it defaults back to 2D clocks occasionally and/or results in worse performance. So in the end I just flashed the card to 1000 mem and left the default core and shaders alone.

Is this normal? Were the G80/G92/G92b etc that desperate for memory bandwidth? I guess they were, Anandtech also found some bottoming out with older versions of the architecture:

http://www.anandtech.com/print/2175

Anyone else noticed any improvements from hitting high shader and core clocks with this older tech?

 

[cusideabelincoln]

Well that is the green version, which runs at lower clocks. The standard memory speed is 1100.

 

[Deanodarlo]

I didn't expect great performance; got the green as it runs cool and quiet, plus a lower wattage as it's in a HTPC. Just wanted other observations as I'd imagine lots of poeple used these cards either now or in the past. This card may use the GTX 285M mobile chip, the G92b-426. Perhaps that makes the difference.

What I was surprised about was how it seems to get no performance increase from higher shaders and core, especially since they are set so low in the first place (738 vs 675, 1836 vs 1456). The only thing that seems to matter is the mem clock! I don't think it's CPU limited, as it's on a core2@ 3.6Ghz.

I was asking as nearly every overclocking thread I read, people push the core and shaders as high as they can, but I'm thinking what's the point? The G92 series seem to be memory bandwidth limited - it was probably held back with even 1200Mhz GDDR3. or perhaps it's another limitation.

Higher core and shaders may help in CUDA and PhysX I suppose?