Mantle to Allow VRAM Stacking in Multi-Card Setups

Yes, stack-able VRAM is almost upon us according to AMD’s own Robert Hallock. Hallock tweeted out earlier this week that the newest addition to the Mantle API would allow VRAM stacking between cards in multi-card setups.

For those of you who don’t know, currently when you take say two 3GB cards and run them in Crossfire/SLI you don’t have 6GBs of usable VRAM, you just have 3. The reason why this is, is because when you are running a multi-card setup, each card is running its own copy of the game individually. The newest rendition of the Mantle API removes the need for each card to run their own copy of the game and therefore their resources can be shared.

This type of functionality will be seen in the future with other APIs like DirectX 12 and OpenGl(glNext) but for now Mantle will be the first to allow it. Although AMD is allowing Khronos access to their Mantle code so it is possible that glNext will be rolling out this functionality along side Mantle.

Now I feel like I should clarify that just because this functionality is available, that does not mean that suddenly all of the games you are playing will be using stacked VRAM. These APIs allow developers more to-the-metal access off the GPUs so stacked memory will only be available in games that the developer optimizes to do so.

21 thoughts on “Mantle to Allow VRAM Stacking in Multi-Card Setups”

  1. No, each card is not “running a copy of the game individually”. With alternate frame rendering the GPUs take turns rendering frames, that’s why they need to have the same data in the cards VRAM.

    This change in DX12/Mantle won’t allow for full 8 GB of VRAM on two 4GB cards at all as there still needs to be sharing for multi-GPU rendering to work. But it should provide for a bit more VRAM to work with.

    1. Explain how it’s useless. Games are getting more and more vram heavy and if you up the resolution that needs more vram.

      This will be very good for those who want to game at above 1080p. Heck Shadow of Mordor and CoD AW both use near 3.5GB VRAM on my 970 at only 1080p (yes I know 3.5GB huehue). If I cranked it up even more which I would if I had a 2nd gpu my vram would be a bottleneck.

      1. Yeah, youre right. If we are talking about resolution and buffers, than this can help. If your memory will be absolutely full on 1080p, than with sfr you can have double resolution. (but i dont understand why do you need mantle for this? It is not a standard property of SFR?). I was talking about resources like 3d models and textures. I believe those still will be on both cards in majority of games.
        EDIT: I can think of game types (static camera for example) when you can have on both card completely different data.

      2. It’s useless because of bandwidth – the only way to get some data from one card to the other is through PCI-E connection and it’s very slow: only 16GB/s (PCI-E 3.0 x16) when current internal VRAM has about 190-230GB/s. If they would add something like SLI-Bridge but with high bandwidth (like 100-150GB/s) – only then this makes sense.

        1. Spatial and temporal locality. Most data should be optimized to stay resident in vram as long as possible frame-to-frame, so saturating the PCI-E bus shouldn’t become much of an issue in the general case.

  2. Skyrim is the only game where people want this…

    I’m sorry but for now this seems marketing talk, because yes developers will also have to put some effort into their games for them to utilize this memory stacking feature. Seeing laughable multi-gpu support and piss poor scaling in several of today’s games, this will happen with only 1 out of 10 AAA titles where you need the grunt.

    Though they should bring this to Mantle in BF4. Mantle is a memory whore and causes troubles for 2gb gpus, and at some settings even 3gb gpus have trouble. However IF amd can stuff this in BF4 I’ll admit I’ll be surprised. Seeing how DICE LA is so open,good, I’m sure they might make it happen.

  3. In other news, AMD is putting finishing touches on 300 series. All brothers who sold the gtx970 with 512mb nvidia pen drive have something new to behold very soon!

  4. This is good news and will help from going into system ram which causes slow down when going over your V-ram limits when using multi gpu’s.

  5. i’ve been around amd long enough to know this won’t affect gcn 1.0 cards and i’m kind of sad because the 3gb on my dual 7970s gets smaller every time i look at my 34″ ultrawide vrams usage on new games

  6. If DX12 will have this feature too, now I REALLY can order my GTX Matrix without worries about future 4k issues due to low VRAM in future games. I almost kept waiting for GTX 8GB news. Now i think it’s not needed anymore, since my next move will be putting it in SLI a year from now.

    It’s not strange that ATI releases this feature (and don’t get me wrong, many thanks ATI!), since they are the only competitor with non-exclusive card models that already have 8 GB VRAM? And….. am i understanding something wrong or it means that a SLI of dual R9 290X will have 16 GB of stacked VRAM?
    o.O

Leave a Reply

Your email address will not be published. Required fields are marked *