How important is video memory for DCS?

How important is video memory for DCS?

I know that it will cause issues if you do not have enough memory, but for example, if you have a video card that has 2GB of video memory, would you see a performance increase in DCS if only its memory got upgraded, let's say... to 4GB?

Depends on what resolution you are running.

I'm at 1920x1080, but what about other resolutions or multi-monitor setups too?

The more pixels the card has to render, the more memory it needs. In my experience, 2GB is absolute minimum for a standard HD display.

Multimonitor setups, yes, you'll want more memory, typically.

But do remember that there's more to video memory than just the amount of it and it's "clock frequency", just as important (sometimes even more so) is the speed of the memory bus. Having extra vRAM in a card doesn't help if the bus isn't fast enough to keep that IO going - but sadly something that many vendors trick you on when they make versions of a GPU with "extra" vRAM.

Gets a bit more complicated since you want actual I/O throughput to do a reasonable metric; there's two variables there - the "width" of the memory bus, and the "speed". Some cards will have a narrower bus, but compensate with a faster one. (Think of it like a highway: you have X amount of lanes, with cars driving at Y speed. X-1 lane can still perform better if it comes with the cars driving faster.)

For 1080p, I'd say 1GB is actually plenty as long as the bus is up to snuff. For more, you'd probably want to look at increasing the memory - depends on the effects used etcetera - though I've used 2-screen setup with DCS on a 1GB card with no issue; though that has been one monitor for DCS and one for windows, so that is a bit different. But don't blindly compare two versions of the same GPU card where the only difference is amount of vRAM, you might get disappointed.

I see.

Another question: If you start with enough memory and add more, will you see improved framerates in any instances?

Lol, you sniped this question. Let's say the card is at 256-Bit.

Only if you were previously memory-starved. An example I would potentially imagine would be a 1GB card being asked to Eyefinity three 1080p displays, though I haven't tested that myself. (And, of course, if the card in question has a crap memory bus, then it still won't really help.)

What I would suggest, in general, is to simply keep in mind what amount of vRAM is on the "reference design" versions of a given graphics card. The bus will be tailored to fit that, and you'll have a guarantee of non-issue. Sometimes you do have considerably more margin from the bus though, but you'd want to read up fairly close on comparative tests and so on to ascertain that.

Also remember that adding a second card (Xfire, SLI) does NOT give you more vRAM to work with, since all cards will "mirror" the same data. (Unless Mantle and Dx12 change this in the near future.)

Only if you were previously memory-starved.

Ok that's what I was getting at. Thanks.

Also remember that adding a second card (Xfire, SLI) does NOT give you more vRAM to work with, since all cards will "mirror" the same data.

What if you Crossfire two cards with different amounts of memory? Which amount is used?

Ok that's what I was getting at. Thanks.

 

 

What if you Crossfire two cards with different amounts of memory? Which amount is used?

The lowest amount of memory is used... So if 2 cards are used and one has 2 Gig and one has 4 Gig you are only going to be able to use 2 Gig in SLI..

EDIT: Same goes for Crossfire..

If you do Crossfire or SLI (my experience), you will see FPS increase. I am running a GTX780 with triple monitors and getting 30FPS on average. My research has shown that I will go up to 40-50fps by adding the second card. With that, I will buy another 780 given that adding a card with more vram will not increase my overall. It basically spreads the workload out over the two cards.

Dale

What kind of research though? This would only possibly be the case if you were previously bottlenecked specifically at the GPU.

If you are bottlenecked anywhere other than specific GPU computation (which omits things like GPU memory/memory bus etc), adding a second card might actually decrease performance, since you'll retain your bottleneck but add a new layer in graphics.

Also note that performance in SLI and Xfire solutions is heavily driver-dependant: in the sense that the vendor (AMD and nVidia) have to put specific support routines for the specific game into their drivers, otherwise you will see less of an increase than what might be benchmarked on Triple-A titles that get a lot of support from the vendor.

If you Crossfire or SLI two cards both at 256-bit will it give you twice as much bandwidth as one 256-bit card? And if so will that mean that at least the mirrored vRAM is being optimized a lot?

If you Crossfire or SLI two cards both at 256-bit will it give you twice as much bandwidth as one 256-bit card? And if so will that mean that at least the mirrored vRAM is being optimized a lot?

No. SLI/Crossfire uses a separate bridge for mirroring, and has one "master" card that handles interaction with the rest of the system.

The memory bus here is between GPU and vRAM, not between GPU and rest-of-system. (EDIT: That is PCI-E territory.)

What happens is this (very simplified, it's more complex than this of course, esp. with the PCI-E interface):

Card0: master

Card1: slave

Card0: gets data/instructions

Card1: receives data/instructions from Card0

Card0: does work

Card1: does work

Card0: receives results from itself and Card1

Card1: sends results to Card0

Then results are out through the framebuffer to the monitors.

Now, this is not necessarily "bad". Drivers can make optimisations that ensure cards handle workload and data optimally for a specific title. It's just one of the technical finicky things that had to be done to make it work (and one of the reasons why there used to be an absolute requirement for the cards to be identical). But if there is no driver optimisations going on, they have to use a fallback mode which would typically be alternate frame rendering, where one card renders one frame, the second the next one, then back to the first one (unless there's three cards). This however requires that there is a buffer of information for each of those frames already prepared and well syncronised, which didn't always work that well (esp. in titles with none or bad driver support), leading to microstutters - where overall FPS would be good, but there would be an interval where a frame gets severely delayed.

Bottom line is that Nvidia is noted as worked very will in SLI mode for this simulations and it will increase the FPS which is the goal in this case I BELIEVE.

Dale

No. SLI/Crossfire uses a separate bridge for mirroring, and has one "master" card that handles interaction with the rest of the system.

 

The memory bus here is ......

Interesting and very helpful. :thumbup:

I got another one ;)

If you're running two different cards in Crossfire which one should you make the Master Card?

Bottom line is that Nvidia is noted as worked very will in SLI mode for this simulations and it will increase the FPS which is the goal in this case I BELIEVE.

 

Dale

I'm trying to get more stable FPS. I get great framerates in open air but it dips a little too much for comfort when I'm in graphic heavy areas with lots of objects or when there are many explosions and such.

The more pixels the card has to render, the more money it needs...

Actually I read like this at first. :)

I'm trying to get more stable FPS. I get great framerates in open air but it dips a little too much for comfort when I'm in graphic heavy areas with lots of objects or when there are many explosions and such.

In that case it isn't just video memory (which you can't upgrade without buying an new card anyway) but the the combination of many things including the CPU and the video card.

Wait for EDGE and hope.

In that case it isn't just video memory (which you can't upgrade without buying an new card anyway) but the the combination of many things including the CPU and the video card.

 

Wait for EDGE and hope.

Yeah I know it's a combo of things. I'm doing Crossfire for the first time, mainly cause EDGE is coming and will take advantage of it. I'm doing it with two different brands of cards so I wanted to know how the different specs interact. Everyone explained it very well. :thumbup:

Something like:

If the first video card is 2 times stronger than the second one.

The first video card will run 2/3 of the monitor panel area.

Something like:

If the first video card is 2 times stronger than the second one.

The first video card will run 2/3 of the monitor panel area.

This makes it sound as if you're saying both cards are rendering the same frame. This is at least typically not the case, and could even reduce performance compared to the "standard" method - the problem is that in this case they would have to partition meshes, effects, postprocessing etcetera several times between themselves even within the same frame. An easier method would be to have (for example) a framebuffer of 3, and have the stronger one do 2 frames while the weaker does 1, then rinse and repeat.

There are tricks that go this route though (esp. with Mantle and potentially Dx12, I hear, where you can obtain similar effects as CFX/SLI without data bridges etc), and I might have missed a specific implementation that gets around the standard "problems" with that route. (Would love links to papers/indepth articles on it in that case! Sadly work has kept me busy enough lately that I haven't been able to read up as much on this stuff as I used to.)

But there are issues present that are not negligible: for example many lighting and shadowing effects require the rendering of the frame to be aware of everything that is in the frame. (And outside of the frame, too...) Having part of the frame be done by an independent unit is then problematic - but I guess could be solved through having the routines be aware of this and have a sort of pre-processing stages that ensures necessary information is passed. (Again, sort of what I think AMD is trying to do with Mantle and Microsoft at least are looking at "enabling" people to do in Dx12.)

Hopefully EDGE comes before the end of this year so I know if I need a new GPU for it. My GTX470 soldiers on with everything else I run but my upgrade window is small.

The rendering frame mode.

This is called "Scissors mode" or "SuperTiling mode"...

http://www.hardwaresecrets.com/article/SLI-vs-CrossFire/391/5

Thank you!

Once again thanks for all the info everybody. I hooked up Crossfire yesterday, and it all went smoothly. It was pretty much plug-and-play!

I did a little performance benchmarking:

--DX 11 FPS w/ one card: about 74

--DX 11 FPS w/ two cards: about 154

Of course those aren't game framerates, but it looks like now I'll be ready for EDGE!

Fingers crossed for a December release!

:D :D :D

The rendering frame mode.

This is called "Scissors mode" or "SuperTiling mode"...

http://www.hardwaresecrets.com/article/SLI-vs-CrossFire/391/5

Thank you!

Cheers, Tiling makes a lot of sense (and I'm reading Scissors and being similar to tiling in "logic", just different in implementation). The logic used in these seems to be similar to what 3D rendering software often does. Seems quite smart indeed if the implementation manages proper load-balancing. (Note that CFX, at the time of publication for that one and until at least fairly recently (late last year), had very high microstutter frequencies on those modes - causing jerky animation in spite of nominally high "framerates". I'm guessing the load-balancer was at fault. Several benchmark sites started recording and displaying frametime results rather than "framerates" or "FPS" due to this, because the "framerates" were very misleading whenever this problem manifests.)