Video Memory width, versus speed. performance comparison

[jiimbob]

I am in the market for a new video card, and am curious about the different specs of ati and nvidia cards.
ati's hd 4870 has far more memory speed at 3.7 ghz gddr5 on some cards, but all their cards have the same 256 bit width. Nvidia offers 448 and 512 bit gddr3 at up to 2500 mhz or so, and I am wondering how these 2 specs effect performance. Nvidia solutions are a little more expensive, but I found a pretty good deal on a gtx 260 ssc overclocked at 675 mhz up from a 576 mhz base clock on the vanilla board. a radeon 4870 is about $50 cheaper with 100mb more memory.
Another option is a gtx 280 at a lower clock rate for $40 more than the ssc gtx 260, around 602 mhz. this would upgrade the shader proccessors to 240 from 216, and memory width to 512 from 448, with gpu and memory speed decreased.
So I'm wondering how these things affect performance. I'v thought, maybe the 4870 has too high a memory speed for its 256 bit width, making it bottlenecked, but I have nothing to base this off.
Thanks for any insight on the subject

 

[TC91]

core/shader clock speeds shouldnt matter too much, comparing apples to oranges there since they are different architectures. The effective memory bandwidth is what you should be looking at for comparison, as the high clock speeds of gddr5 (technically its quad pumped) makes up for its lack of bus width with both the gtx 260 and 4870 offering similar effective memory bandwidth.

 

[jiimbob]

cool, thanks for clearing that up

 

[dflynchimp]

memory bandwidth is calculated by (mem speed)X(mem multiplier)X(bus width)/8

not sure what the 8 is, but it fits because comparing the 4870 with GDDR5 and GTX280 with GDDR3 yields

4870= 900MHz X 4multi X 256bit bus /8 = 115.2GB/s

GTX280= 1107MHz X 2multi X 512bit bus /8 = 141.7GB/s

GTX260= 999MHz X 2 multi X 448bit bus /8 = 111.9GB/s

the quad multi on GDDR5 means the 256bit bus can do what the 512 can do at the same base clcok speeds. If you look at the ratio of the 1107 to 900, you find it's exactly the same as the ratio of 141.7 to 115.2.

From a manufacture standpoint, a 256 bit bus is alot cheaper to make than a 512. The early move to GDDR5 might have cost AMD a good chunk of R&D, but they're making it up with higher margins from the combined cost reductions of a die shrink and a smaller bus.

On the other hand, GDDR3 is very mainstream and cheap, so Nvidia has its own little bit of cost savings by not jumping to GDDR5 early. It looks like for the most part GDDR4 is being skipped over by both companies (minus some exceptions) because it doesn't offer enough perks over GDDR3. The real coup de grace is GDDR5's 4X multi vs 3 and 4's 2x

The numbers above only pertain to memory bandwidth, which isn't the end all indication of actual performance by any stretch of the imagination, but they give some insight into the different strategies

 

[alcoholbob]

1 byte = 8 bits.

 

[dflynchimp]

tyvm

 

[jiimbob]

I was under the impression that gddr3 had a 3x multiplier and gddr5 was only a 4x. so I always though that they just skipped a number for the hell of it, and called it gddr5 instead of gddr4. This is the first i'm hearing of gddr4 being skipped. anyone know for sure about this?

 

[BlueAcolyte]

The GDDR3 uses a 2x multiplier.

AFAIK only the HD 3870 and some speciality ATi cards used GDDR4.

 

[dflynchimp]

Qimonda, the maker of GDDR mem, has intentionally skipped GDDR4 (look it up on wikipedia) because GDDR5 offers so much more benefits. Too bad Qimonda is still doing like shit right now (filed for insolvency end of last year: the German equivalent of bankruptcy)

In general, it's a sad state for the IT sector, with all the spending cuts and layoffs just to break even. AMD in particular got hit huge even taking into account the success in the graphics department.