Memory Bus size and how it effects your VRAM usage

[Kuzma]

Introduction:

You may have seen me post around quite a bit telling people not to buy the 4GB variants of the 760, 770, 660 TI or 680 or the 6GB variant of the 7970; this is due to a little factor called the memory bus size it's hard to explain it on a technical level while still making it easy to understand so I'll simplify it down to a real world analogy.

 

What exactly is the memory bus?

The memory bus is the pathway that your gpu uses to access the memory it has and is generally measured in bits (8 bits = 1 byte :P ) this works together with the memory clock speed to work out exactly how much of the memory can be accessed per second.

 

So how will it effect my graphics card?

Think of the memory as water and the memory bus as a tunnel, if you need more water than your memory bus will let through then you're going to have to wait a while to wait for that extra water to come through. If your graphics card has memory bus designed for 2GB and your add another 2GB then you've added more water without being able to get that extra water through the memory bus.

 

What about memory clock?

The memory clock is like the speed of the water, if you increase the speed of the water enough then you can push more water through the small memory bus ^_^ the issue is however that you need a pretty large speed increase to access double the water in comparison to before.

 

Conclusion

So if you're buying a variant of a graphics card with double the memory then make sure that either the memory clock is increased or you know you'll be able to increase it (gpu boost 2.0) otherwise all that extra memory (and that extra cash) is wasted. Since memory clocks generally aren't very very high a good rule of thumb is 128bits and 1000mhz effective memory clock per gb (this can change to 64bits and 2000mhz and vice versa so make sure you do your math :D to work out if you're going to be able to use all that memory)

 

P.S.

I thought I'd add in my Titan calculations for any of you mathematicians reading this (attatched as a txt) and by my calculation a titan would need exactly double it's effective memory clock speed to access all 6GB of it's memory that or it's only accessing 3GB of it's memory and since we've already seen the titan use more than that leaving me with the conclusion GK110 only needs half the memory clock to access all the GB which to me is a crazy revolutionary advancement O.o

titan calculations.txt

[CyberJesus88]

Wow, great explanation, that helps a lot!

[Kuzma]

Wow, great explanation, that helps a lot!

Thanks and no problem ^_^ I aim to explain.

EDIT: I'm a poet and I didn't even know it

Edited August 9, 2013 by Kuzma

[AMD_Nemesis]

Well Ive officially tested that method in skyrim with a bunch of mods and some how got it up to 3.3 Gbs I believe you are are correct in the sense that bus size effects access rate however I do not believe it effects total capacity of vram usage and if the titan can access more memory than what your calculations say then your theory is invalid its not like its using some high tech super-bus or comes with special vram its a different chip and is meant to be high-end (and overpriced), on very low end cards like a 640 or lower 4gb or something like that I can see the card just stop using more vram since its not even a good card nor built to be a performance king and just cant handle all that stuff. that's just my theory.

Linus should do tests on this on a 7604gb, 7704gb, Titan.

[Bluuz]

Very simple and helpful explanations as always.

 

To add to your metaphor, wouldn't the memory size be more like the width of the tunnel?

[xDLGx]

Very simple and helpful explanations as always.

 

To add to your metaphor, wouldn't the memory size be more like the width of the tunnel?

The memory size is the mass of water trying to get through the tunnel. The bus is the width of the tunnel. The memory clock is the speed of the flow... or could call it water pressure I guess.

[Kuzma]

Well Ive officially tested that method in skyrim with a bunch of mods and some how got it up to 3.3 Gbs I believe you are are correct in the sense that bus size effects access rate however I do not believe it effects total capacity of vram usage and if the titan can access more memory than what your calculations say then your theory is invalid its not like its using some high tech super-bus or comes with special vram its a different chip and is meant to be high-end (and overpriced), on very low end cards like a 640 or lower 4gb or something like that I can see the card just stop using more vram since its not even a good card nor built to be a performance king and just cant handle all that stuff. that's just my theory.

Linus should do tests on this on a 7604gb, 7704gb, Titan.

This isn't a theory btw this is in practise ^_^ look at 660 TI and 680 4GB version benchmarks; also you can't use something you don't have access to... Also as I said it seems that GK110 requires half the effective memory clock to access the same amount of VRAM in which case makes GK110 insanely powerful as they could have a 128bit interface with a 500mhz memory clock for 1GB of vram and it'd run extremely power efficient and be very cool. I think the reason GK110 won't ever be used mainstream is because the yield rates are far too low.

 

LinusTech should do tests on the 760 4gb , 770  4gb and Titan to show if Nvidia have upped the memory clock enough on the 700 series versions of the 660 TI and 680 and if the Titan actually can use all it's VRAM.

 

P.S. Skyrim has a 4GB VRAM limit and you can easily reach it if you try.

[Kilerbomb]

Good info, very enlightening. 

[ExDreamer]

That's why older card have 512bit bus even though it have lower vram. For example gtx 285 have 512bit bus even though it use only 1GB of GDDR3.That is because it have memory clock speed of 648Mhz.

[AMD_Nemesis]

Pretty sure the 770 can use near to all it's VRAM I'll test it for 4 gb in skyrim if I can find more hi res mods.

[leipero]

Good, but maybe you should put it under guides section? .

 

Also, i would love to add one thing, sometimes, wide memory bus is useless. When GPU is not powerful enough to utilize capacity of VRAM, and when availalbe BW is overkill in any scenario, basically you get internal bottleneck in graphic card, by GPU itself. Hope this will contribute to the topic ;).

[Kuzma]

Good, but maybe you should put it under guides section? .

 

Also, i would love to add one thing, sometimes, wide memory bus is useless. When GPU is not powerful enough to utilize capacity of VRAM, and when availalbe BW is overkill in any scenario, basically you get internal bottleneck in graphic card, by GPU itself. Hope this will contribute to the topic ;).

It's not a guide it's information :P if my topic's were in the wrong sections I'm sure they'd have been moved by now.

[lachyman]

People buy 4GB variants when they want to SLI.

[AMD_Nemesis]

People buy 4GB variants when they want to SLI.

or it fits their needs eg: skyrim mods big resolution moniter or they want the card to last a few gens.

[JujuShinobi]

I swear you always make complicated things sound so easy Kuzma. Looking foward to more of your works!

[jeffrey_rodriguez]

this is nice!!! very helpful, what's the bandwith for?

[shadizzle]

hello,

Sorry this thread is old, but I do have a question that surrounds just about all of this.. please pardon my ignorance,

Say you have a titan, knowing its limits with the 384bit buffer in a gaming environment. If the "tunnel" is too small for a gaming environment due to the constant rapid I/O required(visual and processing). How would it vary in say a video editing or CAD application that is design to take advantage of the large vram. The rendering itself can be pretty slow depending on model size and or effects. With that said, wouldn't you now be able to take advantage of the remaining memory since the I/O to it is a different nature? The application itself is written to. Due to the nature being CAD/video editing, wouldn't that now allow full 6gb advantage since you're not "seeing" the effects going on, and the time to process being at a slower rate?

thank you

Edit: grammar

[Kuzma]

hello,

Sorry this thread is old, but I do have a question that surrounds just about all of this.. please pardon my ignorance,

Say you have a titan, knowing its limits with the 384bit buffer in a gaming environment. If the "tunnel" is too small for a gaming environment due to the constant rapid I/O required(visual and processing). How would it vary in say a video editing or CAD application that is design to take advantage of the large vram. The rendering itself can be pretty slow depending on model size and or effects. With that said, wouldn't you now be able to take advantage of the remaining memory since the I/O to it is a different nature? The application itself is written to. Due to the nature being CAD/video editing, wouldn't that now allow full 6gb advantage since you're not "seeing" the effects going on, and the time to process being at a slower rate?

thank you

Edit: grammar

Yes you're correct, in rendering applications you don't need access to all the VRAM at the same time or even during very rapid periods of time meaning you could in theory use a GT640 with 4GB of VRAM for rendering and it'd outperform a GT640 with less VRAM.

[TTesla]

Introduction:

You may have seen me post around quite a bit telling people not to buy the 4GB variants of the 760, 770, 660 TI or 680 or the 6GB variant of the 7970; this is due to a little factor called the memory bus size it's hard to explain it on a technical level while still making it easy to understand so I'll simplify it down to a real world analogy.

What exactly is the memory bus?

The memory bus is the pathway that your gpu uses to access the memory it has and is generally measured in bits (8 bits = 1 byte ) this works together with the memory clock speed to work out exactly how much of the memory can be accessed per second.

So how will it effect my graphics card?

Think of the memory as water and the memory bus as a tunnel, if you need more water than your memory bus will let through then you're going to have to wait a while to wait for that extra water to come through. If your graphics card has memory bus designed for 2GB and your add another 2GB then you've added more water without being able to get that extra water through the memory bus.

What about memory clock?

The memory clock is like the speed of the water, if you increase the speed of the water enough then you can push more water through the small memory bus the issue is however that you need a pretty large speed increase to access double the water in comparison to before.

Conclusion

So if you're buying a variant of a graphics card with double the memory then make sure that either the memory clock is increased or you know you'll be able to increase it (gpu boost 2.0) otherwise all that extra memory (and that extra cash) is wasted. Since memory clocks generally aren't very very high a good rule of thumb is 128bits and 1000mhz effective memory clock per gb (this can change to 64bits and 2000mhz and vice versa so make sure you do your math to work out if you're going to be able to use all that memory)

P.S.

I thought I'd add in my Titan calculations for any of you mathematicians reading this (attatched as a txt) and by my calculation a titan would need exactly double it's effective memory clock speed to access all 6GB of it's memory that or it's only accessing 3GB of it's memory and since we've already seen the titan use more than that leaving me with the conclusion GK110 only needs half the memory clock to access all the GB which to me is a crazy revolutionary advancement

Ok can you please try to explain it on a technical level?

And why are @LinusTech and @Slick excited about a 780 ti 6gb?

That means either this whole this is false (no offense) OR they don't know about this. OR they know and they ALSO know about GK110 ability to be able to gave half the memory clock to access 6gb (most likely) but still, why did nvidia never point out this "REVOLUTIONARY" advancement? Or why do these 4b models EVEN exist? Just a huge scam? Why do more people not know about this since this is very important, and they end up buying 4gb models ANYWAYS! Lastly if Linus and Luke do know of this why have they not done a better job to inform us on this subject?

[xAcid9]

Ok can you please try to explain it on a technical level?

And why are @LinusTech and @Slick excited about a 780 ti 6gb?

That means either this whole this is false (no offense) OR they don't know about this. OR they know and they ALSO know about GK110 ability to be able to gave half the memory clock to access 6gb (most likely) but still, why did nvidia never point out this "REVOLUTIONARY" advancement? Or why do these 4b models EVEN exist? Just a huge scam? Why do more people not know about this since this is very important, and they end up buying 4gb models ANYWAYS! Lastly if Linus and Luke do know of this why have they not done a better job to inform us on this subject?

 

SLI, these 4GB 256bit card exist for those that want to SLI their cards and play on huge reso with multi-monitor configurations.

 

Anyway, this is not new and it is quite common especially in mobile GPU. I still remember back in 2011, got this GT555M with 192bit memory bus and 3GB of VRAM, that card can't even utilize 1GB in any games.

 

[TTesla]

Introduction:

You may have seen me post around quite a bit telling people not to buy the 4GB variants of the 760, 770, 660 TI or 680 or the 6GB variant of the 7970; this is due to a little factor called the memory bus size it's hard to explain it on a technical level while still making it easy to understand so I'll simplify it down to a real world analogy.

 

What exactly is the memory bus?

The memory bus is the pathway that your gpu uses to access the memory it has and is generally measured in bits (8 bits = 1 byte ) this works together with the memory clock speed to work out exactly how much of the memory can be accessed per second.

 

So how will it effect my graphics card?

Think of the memory as water and the memory bus as a tunnel, if you need more water than your memory bus will let through then you're going to have to wait a while to wait for that extra water to come through. If your graphics card has memory bus designed for 2GB and your add another 2GB then you've added more water without being able to get that extra water through the memory bus.

 

What about memory clock?

The memory clock is like the speed of the water, if you increase the speed of the water enough then you can push more water through the small memory bus the issue is however that you need a pretty large speed increase to access double the water in comparison to before.

 

Conclusion

So if you're buying a variant of a graphics card with double the memory then make sure that either the memory clock is increased or you know you'll be able to increase it (gpu boost 2.0) otherwise all that extra memory (and that extra cash) is wasted. Since memory clocks generally aren't very very high a good rule of thumb is 128bits and 1000mhz effective memory clock per gb (this can change to 64bits and 2000mhz and vice versa so make sure you do your math to work out if you're going to be able to use all that memory)

 

P.S.

I thought I'd add in my Titan calculations for any of you mathematicians reading this (attatched as a txt) and by my calculation a titan would need exactly double it's effective memory clock speed to access all 6GB of it's memory that or it's only accessing 3GB of it's memory and since we've already seen the titan use more than that leaving me with the conclusion GK110 only needs half the memory clock to access all the GB which to me is a crazy revolutionary advancement

Ok can you please try to explain it on a technical level?

And why are @LinusTech and @Slick excited about a 780 ti 6gb?

That means either this whole this is false (no offense) OR they don't know about this. OR they know and they ALSO know about GK110 ability to be able to gave half the memory clock to access 6gb (most likely) but still, why did nvidia never point out this "REVOLUTIONARY" advancement? Or why do these 4b models EVEN exist? Just a huge scam? Why do more people not know about this since this is very important, and they end up buying 4gb models ANYWAYS! Lastly if Linus and Luke do know of this why have they not done a better job to inform us on this subject?

One last thing I do remember either Linus or Paul (newegg) saying to get 4gb models if you want to play games even in 1080p with super high res textures (skyrim) that means one of them either has not heard of this (unlikely) or they do not believe it to be correct. So you can use 4gb of ram?

[xAcid9]

One last thing I do remember either Linus or Paul (newegg) saying to get 4gb models if you want to play games even in 1080p with super high res textures (skyrim) that means one of them either has not heard of this (unlikely) or they do not believe it to be correct. So you can use 4gb of ram?

 

Yes, if you want to play Skyrim with tonnes of high res textures mod then 4GB card is the one to get. This is the only game that able to use more than 3GB on that 256 bit bus @1080p. Your FPS will plummet significantly though.

[TTesla]

Yes, if you want to play Skyrim with tonnes of high res textures mod then 4GB card is the one to get. This is the only game that able to use more than 3GB on that 256 bit bus @1080p. Your FPS will plummet significantly though.

BUT if what he said is true then this shouldn't be possible since there is too much water to get through the 256bit tunnel! if this can get through then WHY can't anything else? eh?

[Kuzma]

BUT if what he said is true then this shouldn't be possible since there is too much water to get through the 256bit tunnel! if this can get through then WHY can't anything else? eh?

Yeah, the 4GB makes no difference under any scenario besides compute so xAcid9 is wrong.

[TTesla]

Yeah, the 4GB makes no difference under any scenario besides compute so xAcid9 is wrong.

Why is compute a exception? Doesn't the vram ( water) STILL have to got through the bit tunnel?