How much faster is video card memory compared to system memory?

7331

Limp Gawd
Joined
Feb 28, 2007
Messages
220
Tim Sweeny's comments about virtual texturing (video card using system RAM for textures when the card's RAM is full):

"Virtual texturing eliminates the video memory bottleneck on texture size; whereas in DirectX 9 the size of textures we can use with full performance is limited by the amount of video memory, in DirectX 10 it is only limited by total system memory." Furthermore, Tim predicts that virtual memory will enable a "2X-4X increase in texture usage in games, which will be great for Unreal Engine 3 games, where textures are often authored at very high resolutions like 2048x2048, and then scaled down on lower-end systems to improve performance."

Now, how much of a speed difference is there between the two RAM types? Will a game that requires 600MB frame buffer run just as fast on a GTS 320 as on a GTS 640? Can we depend on Sweeney's comments and go buy a lower memory card now?
 
video card memory is MUCH MUCH MUCH faster than system memory. System ram is on a 128bit bus in dual channel AFAIK, whereas nvidia is pushing 384bit and the r600 is rumored to be 512. This isn't even taking into account how much faster graphics card ram is, pushing close to 1GHz/2GHz DDR whereas system ram has trouble hitting 500MHz/1000DDR. However now that ram system ram is becoming a good bit faster, I can see it being used for textures.
 
It often makes me wonder "why" video ram cant be used as system ram. (like why they dont use the video ram chips to build system ram)
 
It often makes me wonder "why" video ram cant be used as system ram. (like why they dont use the video ram chips to build system ram)

It's the bus that makes it fast rather than the chips themselves.
 
System memory bandwidth is pretty inconsequential if it's being used by the GPU. You can make it as fast as you want, but the amount the GPU can use is limited by the PCIe bus. With a PCIe 32x card, the bandwidth would probably still fall short of your system memory's theoretical maximum. Even if your system bus could handle 100TB/s, you still can't squeeze more than 8GB/s through a PCIe bus.
 
System memory bandwidth is pretty inconsequential if it's being used by the GPU. You can make it as fast as you want, but the amount the GPU can use is limited by the PCIe bus. With a PCIe 32x card, the bandwidth would probably still fall short of your system memory's theoretical maximum. Even if your system bus could handle 100TB/s, you still can't squeeze more than 8GB/s through a PCIe bus.

Exactly. On the card, GDDR is fucking FAST. The GPU has close to 80gb/s bandwidth (G80 card). But it would be impractical for the CPU to access this memory, as the PCIe bus is so limiting (in this case).
 
I agree with you guys, but Sweeny is not someone who talks out of his ass. He programmed Unreal so I guess he knows a thing or two. If he's so excited about this technology, there must be a reason.
 
I agree with you guys, but Sweeny is not someone who talks out of his ass. He programmed Unreal so I guess he knows a thing or two. If he's so excited about this technology, there must be a reason.

Virtual Texturing?

AKA...

Turbocaching or Hypermemory?

This is nothing really new just in software form now.
 
wasn't this the whole buzz back when AGP first came out?

the idea of using system RAM to get around the limit of onboard video ram.

if we use that as an example, you can see how the System RAM with the AGP system really never took off. It was a great idea, at least in theory, but it was just too slow and never used..
 
wasn't this the whole buzz back when AGP first came out?

the idea of using system RAM to get around the limit of onboard video ram.

if we use that as an example, you can see how the System RAM with the AGP system really never took off. It was a great idea, at least in theory, but it was just too slow and never used..

Yup, it was called Direct Memory Execute (DiME), and it was a failure. While you would normally DMA the textures into local graphics memory, and then use them, DME allowed you to use them from main memory.

Sounds great, right?

Except, in a game with lots of textures, your bandwidth gets eaten quickly. Unless you have a texture cache, every single pixel that references a texture must wait for that texture read from main memory. The better solution is to just DMA the textures into it and reuse them without penalty.

In-place texture access is still a pipe-dream, and always will be. Son long as graphics memory remains cheap, nobody will bother with direct execute.
 
I agree with you guys, but Sweeny is not someone who talks out of his ass. He programmed Unreal so I guess he knows a thing or two.
I don't dispute that he knows what he's talking about when it comes to graphics hardware / software, but I just had to bring this up :D .
Tim Sweeney, in a Jan 2000 GameSpy interview, predicted that advancements in CPU technology would make GPUs obsolete by 2006-2007.

Not that I'm holding it against him. Predicting the future is a very tricky business. Just pointing out that even people like Tim Sweeney slip up from time to time.

But all that aside, you can't argue with the numbers. A G80 will have its memory bandwidth cut by a factor of 20 going from an 80GB/s internal bus to a 16x PCIe connection.
You may as well be using the page file :p .
 
Tim Sweeny's comments about virtual texturing (video card using system RAM for textures when the card's RAM is full):

"Virtual texturing eliminates the video memory bottleneck on texture size; whereas in DirectX 9 the size of textures we can use with full performance is limited by the amount of video memory, in DirectX 10 it is only limited by total system memory." Furthermore, Tim predicts that virtual memory will enable a "2X-4X increase in texture usage in games, which will be great for Unreal Engine 3 games, where textures are often authored at very high resolutions like 2048x2048, and then scaled down on lower-end systems to improve performance."

Now, how much of a speed difference is there between the two RAM types? Will a game that requires 600MB frame buffer run just as fast on a GTS 320 as on a GTS 640? Can we depend on Sweeney's comments and go buy a lower memory card now?

Sweeny isn't talking about system memory being adequate speed-wise to allow virtual texturing, he's only saying that using system memory can allow you to use higher resolution textures than you could if you only used that on the video card. At no point does he actually talk about memory speed, just capacity. You can bet that virtual texturing will cause a HUGE performance hit (it's like saying "forget buying RAM, just use the hard drive's page file") and is nowhere near as good as buying a video card with more memory.

Don't run out and buy the 320MB GTS based on what Sweeny said. He discusses what DX10 has made possible, not something a hardcore gamer would want to rely on.
 
As I understand from what he said, he's saying that this feature will let him use 2x-4x bigger textures in next gen games like his Unreal Tournament 3. So it's not just on paper, I think he'll actually use this. We have to wait and see.
 
Thats great that tim sweeny is talking about this,

I hope it is soon to be a reality.

I have 1gb of DDR1 Ram in my system tweaked to extremely low timings and for all my gaming needs its bascially so far, good for full settings on most of the games I play.

However lately I do get a few hitches with stuff like Stalker, etc. No biggie.

I have a 8800gtx 768mb that would make all the difference if it could be utilized for what cant go in system ram.


I was actually talking about this months before Tim Sweeny is, but the only credit I accept for it is the credit that I spread this idea into the consciousness Grid, and tim sweeny picked up on it. Now its sure to materalize and manifest, and thats all I care about. :)
 
System RAM is slower than vRAM thats for sure, but on the other had its still a damn sight faster than loading it from your hard drive...

Also, some of you seem to think he's suggesting that it's used to extend the vRAM, I don't think thats the case at all. I think that if you can cache the textures you know you'll need soon into the system RAM, you can basically swap textures from vRAM to system RAM as and when you need the new textures.

This is possibly something they're going to use to help make the loading in the Unreal3 engine seamless, keep all the next levels art in system RAM and slowly start swapping it out for unused textures in vRAM when the level nears a close.
 
Personally I think they're crazy to try using system memory for using textures. I don't see any reason not to continue with the DX9 style of storing everything in system memory and then loading it into video memory as needed. The performance hit of crossing the PCI-E bus just seems to great to be bothered with.

Here's a few numbers to consider:
-Fastest DDR2 ~10GB/s
-G80 ~86.4GB/s
-R600(w/ 1.4GHz GDDR4) ~179.2GB/s
-PCI-E x16 ~4GB/s

The main thing to keep in mind is that system memory and video memory have different functions. System tends towards lower latency at the cost of bandwidth. Video memory is the exact opposite. It uses higher latency to achieve much greater bandwidth. While GPUs are designed to be good at hiding latency you'd really have to start having an enormous amount of operations in flight to hide the additional latency picked up by crossing the PCI-E bus.
 
It isn't really about the difference in the speed of the memories - it's the difference in bandwidth accessing the memory. I've seen DDD2 800 dual channel bandwidth quoted as 12.8GB/s. To get to it, the video card has to go through the PCIe bus, which has a max theoretical bandwidth of 4GB/s in each direction for x16.

The quoted bandwidth for on-card memory in the 8800GTX is 86.4GB/s. That's a 20x difference.

In reality, it's even worse. Benchmarks using CUDA on the 8800GTX show the actual bandwidth to the GPU from video card memory is around 65-70GB/s. (Not too shabby compared to the max theoretical bandwith.) Actual bandwidth to the 8800GTX from system memory has been measured at only 2GB/s. That's a 30-35x difference.
 
I think that if you can cache the textures you know you'll need soon into the system RAM, you can basically swap textures from vRAM to system RAM as and when you need the new textures.
Isn't that more or less how it's done already? I can't see any other logical way to go about it. I'm pretty sure games don't just crash once the video memory's filled up, so you must be able to swap data in and out on the fly. And where would you be swapping from, if not main memory? The hard drive? :D
 
Back
Top