What to look for when buying a GPU.

[polarbear6]

So a person wants to buy a gpu, what do you recommend looking in it .. ??
memory clock, core clock, shader clock, video ram, memory interface width, model number ... Instead of looking at the benchmarks.

Does having a higher memory interface width(128 bits vs 256 bits) creates any difference, if so, how many fps ??

 

[WelshBloke]

Originally posted by: polarbear6
So a person wants to buy a gpu, what do you recommend looking in it .. ??
memory clock, core clock, shader clock, video ram, memory interface width, model number ... Instead of looking at the benchmarks.

Does having a higher memory interface width(128 bits vs 256 bits) creates any difference, if so, how many fps ??

Whats wrong with benchmarks?

Always better to look at gains in the real world not theoretical ones.

 

[Lonyo]

You have to look at everything in relation to everything else.
The easy way to do it is to look at benchmarks.
The hard way is to work out the memory bus, memory speed, and then calculate memory bandwidth, and then see what theoretical performance the GPU might give by looking at shader quantity, shader speed, ROPs etc, and then get a theoretical maximum performance number.
Then you have to look at the architecture to see what sort of performance you can get as a % of the theoretical power (e.g. ATI currently has 1.4TFlops theoretical vs NV at "only" around 1TFlop, but in actual game situations that isn't how it goes at all so you have to adjust those numbers).

The easy way, the most sensible way, the clearest way, and the most accurate way... is to look at the benchmarks.

One card may have a 128-bit memory bus, and another a 256-bit memory bus, but that memory bus is meaningless without knowing what type of memory is hooked up to it, and what speed the memory hooked up to it is running at. Then there are game questions, resolution questions, added effects questions, detail setting questions.
Best way... is benchmarks.
The HD4830 has a 256-bit memory bus and 640 shaders with x clock speed.
The HD4770 has a 128-bit memory bus and 640 shaders with x+ clock speed (something higher).
The HD4770 is faster. The 256-bit memory bus is worth negative FPS.

Most of the time it's worth something, kind of, but it's useless to consider any aspect in isolation, and easier to use benchmarks.

 

[polarbear6]

Originally posted by: Lonyo
You have to look at everything in relation to everything else.
The easy way to do it is to look at benchmarks.
The hard way is to work out the memory bus, memory speed, and then calculate memory bandwidth, and then see what theoretical performance the GPU might give by looking at shader quantity, shader speed, ROPs etc, and then get a theoretical maximum performance number.
Then you have to look at the architecture to see what sort of performance you can get as a % of the theoretical power (e.g. ATI currently has 1.4TFlops theoretical vs NV at "only" around 1TFlop, but in actual game situations that isn't how it goes at all so you have to adjust those numbers).

The easy way, the most sensible way, the clearest way, and the most accurate way... is to look at the benchmarks.

One card may have a 128-bit memory bus, and another a 256-bit memory bus, but that memory bus is meaningless without knowing what type of memory is hooked up to it, and what speed the memory hooked up to it is running at. Then there are game questions, resolution questions, added effects questions, detail setting questions.
Best way... is benchmarks.
The HD4830 has a 256-bit memory bus and 640 shaders with x clock speed.
The HD4770 has a 128-bit memory bus and 640 shaders with x+ clock speed (something higher).
The HD4770 is faster. The 256-bit memory bus is worth negative FPS.

Most of the time it's worth something, kind of, but it's useless to consider any aspect in isolation, and easier to use benchmarks.

I needed to conform this to one of my friend.

 

[crisium]

Benchmarks are everything: they show you the strengths and weaknesses of your card not only on its own merits but vis-à-vis other cards. Anything else is just worthless bunk. Bits and MBs only relevant within the same card, ie. they used to make both 64-bit and 128-bit versions of old cards so you'd always want to get 128-bit. But it's COMPLETELY WORTHLESS FOR SEPARATE CARDS (unless you'd rather have a 512-bit 2900XT then 128-bit 4770... which would make you an idiot). Likewise 1GB cards are often faster than 512MB... but not always and especially not always when comparing different cards. Hence the need for benchmarks to find out the situations when the extra memory is vital and decide if it's worth the money for you.

This more or less is also handy if you don't want to hunt for benchmarks:

http://www.tomshardware.com/re...phics-card,2404-7.html

Within each group there can sometimes be a decent performance difference, and also going from one tier to another is never the same performance increase. But it's good for a general feel of how fast a card is and then you can google for benchmarks between competing cards.

The best thing to look for is your budget and then have the people here recommend what to buy. Clue: they're not gonna say "for $150 budget I recommend a 256-bit card with 1GB of memory". They're gonna say get a Radeon 4870.

 

[yacoub]

But, on the other hand, if given two cards running the same architecture, one with a 128-bit bus and sky-high clockspeeds and the other with a 256-bit bus and lower clockspeeds, and they perform about equal in benchmarks, i'd take the wider bus lower speed card every time. It requires less power, puts out less heat, and the fan won't have to ramp up as high or as often, and yet gives you the same performance as the higher-clocked card.

Bus width is, in a way, analogous to engine displacement in cars. One car that needs to scream up to 9,000 rpm to make the same performance numbers out of a smaller engine compared to another car that can put those numbers out at 5,500 rpm with a larger engine, will be stressed much harder to do so. It'll be louder, run hotter, usually consume at least as much fuel to do so, etc. Likewise, a card that has to be clocked to hell just to get it to perform as well with a narrower bus, is not as appealing as a wider bus, calmer card.

So when you have two cards based on the same core, and one has a wider bus and lower clockspeed, and another has a narrower bus and higher clockspeed to make up for it, and they perform about the same, go with the wider bus, cooler, quieter card.

Just sayin'.

 

[dguy6789]

All in all, individual specifications aren't important. It's more important to pay attention to card models and look at reviews. Cards with less fill rate can beat cards with more, cards with less memory bandwidth can beat cards with more.

 

[crisium]

One thing to add is that a lot of manufacturers make factory overclocked versions of the same card. In that case... then yes core clock, shader clock, memory clock make a difference. You can get factory overclocked GTX 275's that encroach on GTX 285 territory. But since you can OC most cards anyway, it's only really important if you either don't want to overclock or read reviews that confirm that one card OCs better than another. But outside of the same cards it's not too useful. Radeon cards usually have higher clock speeds than Nvidia cards but they aren't always faster... that's just the nature of the architecture.

 

[betasub]

^ All good things to consider.

And then there is also power requirements (PSU), driver support, & price.

 

[Lonyo]

Originally posted by: yacoub
But, on the other hand, if given two cards running the same architecture, one with a 128-bit bus and sky-high clockspeeds and the other with a 256-bit bus and lower clockspeeds, and they perform about equal in benchmarks, i'd take the wider bus lower speed card every time. It requires less power, puts out less heat, and the fan won't have to ramp up as high or as often, and yet gives you the same performance as the higher-clocked card.

Bus width is, in a way, analogous to engine displacement in cars. One car that needs to scream up to 9,000 rpm to make the same performance numbers out of a smaller engine compared to another car that can put those numbers out at 5,500 rpm with a larger engine, will be stressed much harder to do so. It'll be louder, run hotter, usually consume at least as much fuel to do so, etc. Likewise, a card that has to be clocked to hell just to get it to perform as well with a narrower bus, is not as appealing as a wider bus, calmer card.

So when you have two cards based on the same core, and one has a wider bus and lower clockspeed, and another has a narrower bus and higher clockspeed to make up for it, and they perform about the same, go with the wider bus, cooler, quieter card.

Just sayin'.

Except when it's the HD4770 vs HD4830, mainly due to different RAM being used, which you ignore and just talk about clockspeeds.
This is why the whole issue is stupid, because bus width in and of itself is meaningless without lots of other elements being taken into consideration.

 

[videogames101]

Architecture.

 

[kalrith]

When the 6600 GT came out, the specs on paper showed that it didn't come close to standing up to the 9800 Pro. However, real-world benchmarks showed that it not only stood up to the 9800 Pro but it bested it in many games.

Specs are important when they translate into real-world performance (i.e. this certain game is a memory hog and needs 512MB to really shine), but looking at specs without looking at real-world performance gains from those specs is pointless.