VRAM capacity vs VRAM bus width

[Whach]

Hi all!

I am curious about VRAM capacity vs bus width.

As I sure everyone is aware, nvidia's new Kepler architecture is rumored to have a 512 bit VRAM bus width with 2GB capacities (granted, all speculation at this point) vs the 384 bit AMD have implemented.

I was curious to know if my interpretation regarding VRAM is right/wrong.

Does increasing the bus width allow for smaller capacities while maintaining the same kind performance because data is moving in and out of VRAM at a quicker rate, despite have less VRAM?

In simpler terms, I have imagined VRAM capacity to be a bucket of sorts, and bus width to be a funnel?

Any clarification is appreciated!

 

[Red Falcon]

It depends on the clock speed of the memory itself.

Technically you could have a 64-bit memory bus, but if clocked at 16GHz, it would have more bandwidth than a 512-bit bus clocked at 1GHz.

In simpler terms, I have imagined VRAM capacity to be a bucket of sorts, and bus width to be a funnel?

Correct, but clock speed is the throughput of that width.

 

[Forceman]

Think of bus width as how wide/deep the river is, and bus speed as how fast the water is flowing. So you can have narrow bus widths with high speed RAM, or wide bus widths with slower speed RAM - and get the same throughput with both. VRAM size is not directly connected to bus width, although the width does drive how much VRAM is easily usable - 384-bit is good for 1.5 or 3GB - 79xx and GTX 580, while 256-bit is good for 1 and 2GB (6950 and 6970).

 

[Whach]

Think of bus width as how wide/deep the river is, and bus speed as how fast the water is flowing. So you can have narrow bus widths with high speed RAM, or wide bus widths with slower speed RAM - and get the same throughput with both. VRAM size is not directly connected to bus width, although the width does drive how much VRAM is easily usable - 384-bit is good for 1.5 or 3GB - 79xx and GTX 580, while 256-bit is good for 1 and 2GB (6950 and 6970).

Gotcha, so say VRAM frequency/speed is a constant, one capacity is 3GB @ 384bit wide, another is 2GB @ 512bit wide, would they be comparable? what would need to change to make them roughly equivalent? (vram size or speed?)

Thanks for the info so far all!

 

[Forceman]

Gotcha, so say VRAM frequency/speed is a constant, one capacity is 3GB @ 384bit wide, another is 2GB @ 512bit wide, would they be comparable? what would need to change to make them roughly equivalent? (vram size or speed?)

Thanks for the info so far all!

Size and speed are really independent. Fast VRAM (high bandwidth) is good because it keeps the GPU fed with data and working as fast as possible (you don't want to be bottlenecked by VRAM speed). More VRAM really only helps in situations where you need a lot (because you want the data to be in VRAM instead of system RAM, which is much slower). So at high resolutions or with high levels of AA, you need more VRAM to keep the system from having to go to main system RAM for the data.

They are kind of tied together in that you don't really need a lot of VRAM on a slow card, and you also don't really need a lot of bandwidth. Higher performance cards need both faster and more VRAM, since they can effectively run the higher resolutions.

It really only comes into play when you start running multiple cards (AMD and Nvidia design the card's VRAM speed and capacity based on its performance). So if you had a pair of 1 GB 6950s, say, the cards together would be fast enough to run high resolutions, but they might nott have enough VRAM to support it, and performance will sometimes plummet because the cards have to go to system RAM for the data.

 

[XacTactX]

This is kind-of OT but if you want to really understand how fast a GPU's memory is, always consider the speed of the memory in bandwidth. The GTX 570 "only" has RAM running at 950 MHz, but since it's on a 320 bit bus, it's bandwidth of 152 GB/s is comparable to the HD 6950, which runs at 1250 MHz, with 160 GB of bandwidth. To sum it up: memory speed and bit depth boil down to memory bandwidth.

Like Forceman said, memory speed and bandwidth aren't interchangeable, for the most part. If you use Afterburner and monitor your memory usage with an nVidia GPU, you will see that memory usage is fairly constant. I doubt that having 1GB of RAM that is 2x as fast will be the same as 2GB of RAM that is 1x as fast. But I could be wrong. I have been before.

 

[LuminaryJanitor]

Does increasing the bus width allow for smaller capacities while maintaining the same kind performance because data is moving in and out of VRAM at a quicker rate, despite have less VRAM?

VRAM is either full or it isn't. All else being equal, there's no difference between 2GB and 10TB when you're only using 1GB of it.

And when you hit your capacity, it's not a graceful decline in framerates like you'd see when you saturate some other component. You get a small hitch whenever you try to access something which isn't there, and these pile up very quickly; there might only be a few dozen MB between butter-smooth and unplayable stuttering.

 

[Whach]

VRAM is either full or it isn't. All else being equal, there's no difference between 2GB and 10TB when you're only using 1GB of it.

And when you hit your capacity, it's not a graceful decline in framerates like you'd see when you saturate some other component. You get a small hitch whenever you try to access something which isn't there, and these pile up very quickly; there might only be a few dozen MB between butter-smooth and unplayable stuttering.

I know what you mean. I had gtx 570's that hit the vram wall in BF3. GPU's were capable, vram capacity was not.

But I still wonder why nvidia would opt for a (rumored) lower vram capacity - hence my curiosity regarding speed vs capacity. Maybe 2GB is plenty, but I am weary for how long that may be the case. Maybe some AIB will make a 4GB variant, like the 3GB 580's. Here's hoping =)

 

[Forceman]

But I still wonder why nvidia would opt for a (rumored) lower vram capacity

Hard to speculate on rumors of course, but it probably has more to do with the design choices they made for the card. If the 660 part is a 256-bit bus, as rumored, then the easy VRAM choices are either 2GB or 4GB, and given the cost of high-speed VRAM the 4GB card would just be too expensive for the market spot it occupies based on performance. I would expect that a high-end part using a 384-bit bus would have 3GB default (since it would either be 1.5GB or 3GB). If they go to a 512-bit bus then you are back to 2GB or 4GB, so that would be an interesting situation where you might have something like the 6950, where both capacities are offered.

 

[bigdogchris]

VRAM size is not directly connected to bus width, although the width does drive how much VRAM is easily usable - 384-bit is good for 1.5 or 3GB - 79xx and GTX 580, while 256-bit is good for 1 and 2GB (6950 and 6970).

I've read this before elsewhere but still don't understand. How does bus width mean more memory can be used?

 

[JMccovery]

I've read this before elsewhere but still don't understand. How does bus width mean more memory can be used?

Here's a small refresher course on graphics memory:

Current graphics memory chips (GDDR3, GDDR5, DDR3) come in (mostly) 32-bit bus widths (right now, no one is making 64-bit GDDR3/5), as a result you have to use multiple ICs;
For:
128-bit, you need at least 4 chips
192-bit, you need at least 6
256-bit, at least 8
384-bit, at least 12
512-bit, at least 16

GDDR3/5 ICs come in capacities ranging from 256mb (32MB) up to 2gb (256MB)

For example, a 256-bit memory bus will yield (mb = megabits, gb = gigabits:
8x256mb = 256MB
8x512mb = 512MB
8x1gb = 1GB
8x2gb = 2GB

The Radeon 7970 and 7950 use 12 2 gigabit GDDR5 chips for 3GB.

Hope this helps.

EDIT: I forgot to add that a few years ago GDDR3 and GDDR4 chips came in 64-bit widths, but they were more expensive, and in lesser quantities than the 32-bit chips.

 

[Red Falcon]

^ Best description I've read of this yet.

 

[dcmo]

Sorry about resurrecting an older post, but I still have some questions about the factors that go into determining the quality of a card and this seemed like a good thread to ask them. Now, it seems to me memory bandwidth can't be the only factor; why then would one need to overclock the core if it were.

Say I have two cards, a Nvdia and an AMD, or vice versa. One of them has more memory bandwidth, but what else should I be looking at to determine which of the cards is actually the better one, or rather more capable of putting out more FPS's.

Sorry about what may be noob questions, just trying to learn. Thanks.

 

[wonderfield]

The best way is to compare benchmarks. GPU architectures vary too wildly between vendors, and between different card generations from the same vendor, to make comparing specifications particularly useful.

 

[dcmo]

Well, I guess you could say I have a OCD complex. Knowing the results are fine, but I would like to know the why's behind it. I would think memory bandwidth and the core clock go hand in hand; I'm not sure of the other factors that play in to it though.

 

[wonderfield]

Memory bandwidth; memory latency; shader unit clock; shader unit configuration; shader unit count; ability to do single instructions on multiple data; efficiency of the scheduler; quality of the driver; number of ROPs...

It's sometimes very difficult to look at a result on one piece of hardware and the result of the other and to be able to precisely pinpoint where the advantage stems from, especially as shader complexity and the number of passes per fragment/per scene increases.

 

[Red Falcon]

Yes, if you are comparing, make sure to compare NVIDIA->NVIDIA or AMD->AMD on hardware.
If you are comparing performance, then look at game performance results and reviews.

gpureview.com is also a great resource for GPU, IGP, and IGC information.