GDDR2 vs GDDR3 videocard memory

[Arcmanov]

Can someone explain how the 7600 GS, with GDDR2 memory could actually be quite a ways faster than a 6800 GS with GDDR3 memory?  I would have thought that the GDDR3 memory would be faster, and by extension, make that videocard faster (notwithstanding the default memory frequencies of both cards).

Also, what effect does the amount of RAM on a video-card (128/256/512 Mb)  REALLY have on performance?

Highly technical explanations are most welcome.

[Spazosaurus]

Then how does a 6800 with 700mhz DDR1 beat a 6600 with 1GH GDDR3? Its not all about the memory. Its the pixel pipelines, vertex shaders and not to mention the more advanced GPU that is the 7 series. More memory means that textures can be stored in an uncompressed format. Small amounts of memory will force tectures to be compressed therefore decreasing image qualty. Not all cards can take advantage of 256 and 512 though. Only the more powerful ones in the 6 and 7 series for NV and xX00 and x1XXX for ati.

Not very technical but this is my general understanding of the situation.

[TriniXaeno]

DDR2 and memory in general has gotten pretty cheap. You'll find copious amounts slapped on cards for selling points. Marketing more than anything else.

FX 5200 cards with 256MB memory and such.

6600s with DDR2/3

but the underlying hardware of the chip will tell the true story of whether that RAM can even be taxed.

I did a pretty comprehensive study on it a few months back. Core speed, mem speed, memory interface, memory type and number of pixel pipelines all add up to determine the real world speed of the card. The results were pretty consistent.

another point to consider, is that the higher end ram gives you more headroom for overclocking. So though it won't make a difference at stock levels, you can get a lot more out of a card with DDR3 memory than say regular DDR.

[TrinireturnofGamez]

 the most important factors in performance are 1st. Core speed , or the fill rate, you can find by multiplying the number of pipelines ( 12 for a 7600) by the clock (560 million hertz for the GT) = 6720 million pixels per second , which is more than 50% faster than a 6600gt.'s 4000 megapixels.
      Memory speed is the 2nd most important factor , this depends on both the MHZ of the memory ,and the memory controller on the card,  While a 7600gt has more clock than some 7900gts, the 7900gt has a 256bit memory controller vs a 128bit on the 7600 , doubling the memory bandwidth... so it still is faster..
   The amt . of memory matters  too.. but 256 is usualy more than enough in this day and age , with PCI-E you can swap video memory to the main memory fast without too much of a performance hit, and few games actually use that much vid mem.. 
       The difference between GDDR2 and GDDR3 is frequency potential , GDDR3 starts at 900mhz and can go up to 1600mhz , GDDR2 starts at around 600mhz and goes up to about 1000mhz ... s o a card with GDDR3 and the same 128bit memory will be faster than one with GDDR2

[TriniXaeno]

yeah,

and the memory on vid cards is used for textures. The amount of vid memory you need depends on the resolution of your textures.

there is no appreciable graphics quality increase after a certain point, which is why few games call for more than 256MB. The ones that do...are just being ridiculous at the moment.

In fact a 128MB card would be fine for most games at standard res textures. The higher memory mainly comes into play with AA. Since anti aliasing works by super sampling a texture, it pretty much means you are running at a much higher text res than your game resolution would suggest.

For example, a 128x128 texture would be super sampled and interpolated at 256x256 to give you that smoother finish.

It's a great effect but uses a brute force method to get the results. The extra memory comes into play to store those bloated textures and keep things running smoothly when you crank that sucker up.

[Arcmanov]

Niceness.  So fundamentally the main difference between those cards is the complexity of the GPU, right?

I waitin on W1ntry to 'bless up' the thread now.
Hit me wit some tech dey............

[Beomagi]

supersampling applies to the final image - texture multisampling is new and is done through multisampling - adaptive resampling and transparency aa are texture multisampling techniques introduced with the 7000 nvidia series, and the x1k.

SIS's cards use supersampling, but it's very bandwidth intensive - impeccable looks though. ATI and Nvidia use multisampling - i.e. pixel color isn't the color at one point, but for 4x, the average color amoung 4 very close points.

High res uses the memory if you're buffering - triple buffering a 2560*1600 screen with 32 bit color means 2.5*1.6*4 bytes per pixel * 3 images reserved for memory (almost 50 MB)

Baego is right in that 128MB is still fine for most. Only a few games react better with 512MB memory vs 256MB. even then the difference is almost unnoticable - e.g. quake 4/doom3. wait, scratch that, it IS unnoticable. I have tried and i just can't tell the difference in textures between the 2 top choices. (more of a design issue than hardware).

for your original question, its a matter of balance. now the 128 bit interface halves the bandwidth of the 1400MHz memory in the 7600gt, putting the memory BANDWIDTH 30% under the memory in the 6800gs, but the 7600gt's pipes are faster. Now, the memory in the 7600gt is clocked higher - while it wont help throughput as it's memory is half the bit interface, it can help the shader instructions performed in the memory - these shader programs wont use much memory at all - it's tiny enough to ignore, but the performance is probably higher for the higher clocked memory as it wont need the bandwidth of the 256bit interface. The creaters can test the cards to see how memory is being used - if they deem that most bandwidth is unneeded, they cut it, and raise the core. Similar balance issue when you consider ati's x1900 - they decided to cut down on texture units and push pixel shaders. This was the opposite of the 5900xt from nvidia. The games then were balanced differently - more textures. So the 5900's were 4 pipe cards, with 2 texture units per pipe and one shader.