GFlops vs memory bandwidth etc

[islandboy77]

Hi all. I've spent the last week, on and off, trying to get a good handle on the relationship between the following aspects of GPUs:
- Memory Bandwidth (GB/sec)
- GFLOPS
- Pixel Fill Rate (MPixels/sec)
- Texture Fill Rate
The reason behind the research is to get a firm handle on whether the video card I'm looking at buying (and those I buy in the future) is really better than other models I'm looking at. For example, on the GPU Review website, when I compare the nVidia GTX 460 against the GTX 560, the 560 beats the 460 in all area except the GFLOPS. And since the 560 is a "younger card" and potentially about the same price second hand, I'm leaning in that direction.
I found some "real world" benchmarks on the Passmark site that indicate that the 560 is capable of scoring about 12% faster than the 460, but when I checked the "Last 5 baselines" sections for both cards, the difference was much less (mostly due to one baseline for the 560 being significantly lower - not sure what to make of that).
I'm currently on a Radeon HD 5770 1GB, but it is struggling a bit on my Core 2 Duo E6850 3GHz system with 4GB DDR2 800 RAM running W7 32-bit playing BF2 / AIX2. Both the 460 & the 560 appear to be significantly better performers than the 5770, and it appears that the BF2 game has been written with a "bias" towards nVidia cards / shaders / drivers etc.
So I don't want this to devolve into an AMD vs nVidia war - I'm actually pro-AMD, but have decided that if the game I play is going to run better on nVidia, so be it.
I'd appreciate thoughts, comments, links and such to help me sort through the technical aspects of comparing the 4 categories of information I listed above.
Cheers

 

[islandboy77]

Bump. Can no one shed any light on this for me?

 

[Sunius]

Hey. Don't check passmark - that is probably the worst benchmark you can use to check for performance differences. Also, specs mean little and it's impossible to check which card is better from them alone. Lastly, I'd advice you to check benchmarks at sites like www.techpowerup.com or www.tomshardware.co.uk and you should be able to compare the cards there.

 

[islandboy77]

Thanks for your post. I realise that using just the tech specs alone isn't ideal, but it does give me a starting point as to what the basic reference performance is. I've seen / used a few cards over the years, and have yet to use one that doesn't at least meet the reference specs performance levels.

I would still very much like to get a handle on how the 4 things I mentioned correlate to performance, as it's useful when browsing specs to know which cards to look into further and which to ignore. Also, I'd just like to know how those 4 things work out for my own personal interest and understanding.

Anyone else able to shed any light on the interaction / importance of the relationship between Memory Bandwidth (GB/sec), GFLOPS, Pixel Fill Rate (MPixels/sec) and Texture Fill Rate?

Thanks.

 

[Sunius]

Thing is, it's too dependent on the architecture. One cards benefit from one thing, others - from the other.

 

[islandboy77]

Thanks again for your reply. Ok, I can see that. Still, I'm very interested to know what the relationship is between those 4 things. What does each one bring in terms of performance? How do the values of each thing directly effect performance of rendering gaming imagery? In know that the GPU Review comparisons between cards is simplistic, but, again, it would be nice to be able to say something like "Well, if one card is strong in Memory Bandwidth (GB/sec), Pixel Fill Rate (MPixels/sec) and Texture Fill Rate, but is weaker in GFLOPS, on the whole it's still ahead of a card that is the other way around and would be worth investigating."

Many times I've been trying to get a handle on older cards (that I can easily get second hand for reasonable amounts) vs new (or much newer) cards that are much more expensive. The thing is, I often find that of the 4 main features I take are pretty essential for good performance, it's surprising how often older cards will "beat" newer cards in 3 out of the 4 categories. Case in point: The GTX 460's & 470's still outperform a fair portion of the 500 & 600 series. Granted, this is only "on paper". But if I can see that a 2-year-old card is, on paper, beating a brand new or near-new card, and is cheaper into the bargain, I'm going to look more seriously into that. At the same time, if the differences aren't great between an older card and a new(er) one, and the price difference isn't too great, I'm going to look harder at the new(er) one(s) since they will also confer an amount of efficiency (potentially) due to their younger technology that isn't available via the older ones.

But having said all this, I want to be able to properly evaluate and weigh up in my mind how much weight to lend to each of the 4 facets of video card specs so that I don't spend wasted hours looking through many cards: in other words, I'm busy enough in life as it is without trolling needlessly through, say 20 cards when, by understanding the technical aspects better, I might be able to pare those 20 cards down to 5 or 6.

As a clarifier, I'm not looking for a link through to "white papers" - I've found a few of those already, and they're too complex / wordy, and are just too time-consuming to grind through - and I'm not THAT interested in THAT degree of information.

 

[islandboy77]

Bump

 

[Anik8]

well ohkay... the GFLOPS is the horsepower of the gpu that enables the gpu to work on ROPs(pixel units) and texture units and the resultant fill rates is a measure of how many such units are present,how many cycles they use the processing power in,how many parallel raster engines they are connected to.In this way the signals to the gpu are converted into graphics(pixels and textures) and let to display on the monitor.
The rate at which the gpu recieves/reads/stores the signals is called bandwidth(Gb/s).