How to maximize a GPU with very limited memory bandwidth?

[cbn]

Here is the card in question:

http://www.newegg.com/Product/Produc...&SID=u00000687

480 stream processors @ 650 Mhz.

.....but then it has 128 bit DDR3 running @ 1066.

That seems like quite an unbalanced ratio of GPU core power to memory bandwidth.

Any particular usage scenarios for light gaming that would allow for maximized GPU core use in such a memory bandwidth video card?

Examples:

Does turning down resolution reduce stress on memory bandwidth more than turning down various detail settings (At a relatively higher resolution)?

Any particular IQ settings that are notoriously hard on memory bandwidth, but yield very little in improved visuals?

P.S. A few years ago, I remember reading that tessellation was supposed to be the cure for limited memory bandwidth scenarios, but I haven't really seen much written since then.

 

[ShintaiDK]

Its not unbalanced. Its not exactly a speedster GPU.

 

[cbn]

Its not unbalanced. Its not exactly a speedster GPU.

In the past, I know cards like the HD5670 (400 stream processors @ 775 Mhz ) benefited greatly from GDDR5 compared to the HD5570 (400 stream processors @ 650 Mhz) with 128 bit DDR3 @ 1600. (The gains were larger than the 19% increase in GPU core clock)

Example:

HD5670: 47.8 FPS
HD5570: 28.7 FPS

That is a 67% increase in frame rate for only a 19% increase in GPU clock.

=========================================================================

The card I linked has 480 stream processors @ 650 Mhz and the bandwidth is even lower at 128 bit 1066 DDR3.

 

[Erenhardt]

Lower resolution and texture quality. Overclock memory first, core second. Then look for better card

 

[toyota]

Its not unbalanced. Its not exactly a speedster GPU.

wrong. it is a VERY unbalanced card. 1066 effective memory completely butchers the card compared to the 4000 memory on other 6570 models.

 

[ShintaiDK]

wrong. it is a VERY unbalanced card. 1066 effective memory completely butchers the card compared to the 4000 memory on other 6570 models.

4000? Are you sure you dont refer to the GDDR3 cards with 1800Mhz effective?

http://www.newegg.com/Product/Produc...82E16814150674

Or do they have 3 variants?

 

[ShintaiDK]

In the past, I know cards like the HD5670 (400 stream processors @ 775 Mhz ) benefited greatly from GDDR5 compared to the HD5570 (400 stream processors @ 650 Mhz) with 128 bit DDR3 @ 1600. (The gains were larger than the 19% increase in GPU core clock)

Example:

HD5670: 47.8 FPS
HD5570: 28.7 FPS

That is a 67% increase in frame rate for only a 19% increase in GPU clock.

=========================================================================

The card I linked has 480 stream processors @ 650 Mhz and the bandwidth is even lower at 128 bit 1066 DDR3.

AA for example is very memory heavy.

But remember the 5570 uses a 400:20:8 config, and the 6570 480:24:8. Its roughly what, 20% difference after accounting for different config.

 

[cbn]

4000? Are you sure you dont refer to the GDDR3 cards with 1800Mhz effective?

http://www.newegg.com/Product/Produc...82E16814150674

Or do they have 3 variants?

Anandtech tested HD6570 with GDDR5---> http://www.anandtech.com/show/4278/amds-radeon-hd-6670-radeon-hd-6570

It looks like it came out roughly equal to HD5670. (Makes sense considering they have the same bandwidth and 400 stream processors @ 775 Mhz is about the same as 480 stream processors @ 650 Mhz.)

 

[DominionSeraph]

Its not unbalanced. Its not exactly a speedster GPU.

The 6570's core is in the same range as an 8800GT.

8800GT: 57.6GB/s
GT 240 GDDR5: 54.4GB/s
3870: 72GB/s
9600GT: 57.6GB/s
5670: 64GB/s
4850: 63.6GB/s

This thing: 17GB/s

They didn't give those cards that bandwidth for nothing.

The GT 240 review shows the difference between its GDDR5 version at 54.4GB and the DDR3's 25.2GB/s:
http://www.anandtech.com/show/2906/5

It's not always horrible, but there's always a difference. The card linked has only 68% of the GT 240 DDR3's bandwidth.

 

[cbn]

Lower resolution and texture quality. Overclock memory first, core second. Then look for better card

Thanks. I did not know texture quality used a lot of bandwidth.

P.S. While looking through the Anandtech HD5570 review, link here, it looks like the gap (among various games) between the low bandwidth HD5570 and high bandwidth HD5670 is not decreasing with lower resolution and removal of AA.

In fact, in Resident Evil 5 the gap between the low bandwidth HD5570 and the high bandwidth HD5670 actually increases when the resolution drops to 1280 x 1024 and the AA is removed.:

HD5670 starts off 67% faster than HD5570 using 1680 x 1050, 4x AA and Maximum Quality settings.

With resolution lowered to 1280 x 1024 and AA removed (but Maximum Quality settings kept the same), the HD5670 is now 97% faster than HD5570.

So yeah, something about detail setting (texture quality, etc) seems like a stronger factor here than resolution or AA.

At this point, I wonder what would have happened if Quality Settings were lowered from Maximum to Low? It would have been interesting to see if that gap between the low bandwidth HD5570 and the high bandwidth HD5670 finally narrowed.

 

[toyota]

texture quality as in the texture resolution? that has nothing to do with memory bandwidth. if you have enough vram then there will be basically zero performance difference between low or very high textures.