Shouldn't DX12 also offer performance gains in multi GPU situations, similar to CPU gains?

[TheMorningStar]

With DX12 offering native support for multi card solutions, on top of the lower level API (unsure if this also sees a benefit from lower level APIs) I'd expect some great gaines here also. There isn't really that much of a buzz about it at the moment with the CPU side getting most of the attention but I've seen some interesting stuff. Things we have looked at before like Lucid (I think it was) that looked at using multiple unmatched cards even from different manufacturers. APUs working alongside dedicated cards for small yet notable performance increases, again reminiscent of Lucid, probably nothing will come of multi manufacturers, because politics but I still think it's all really interesting.

Any articles looking at DX12 multi gpu scaling? I'd be interested to look them over, even if theoretical. I don't expect gains to be instant with everything else the industry has to play around with currently but I like to note the possibilities

 

[Calin Banc]

Like in Mantle, scaling will be better, because the CPU isn't a bottleneck anymore or at least not as prominent as before.

 

[Rroff]

Don't think we'll see immediate gains, developers have a bit more in the way of options for maximising utilisation of available rendering options and/or working at a lower level to avoid compatibility issues but ultimately there are limits both on how you can utilise multiple GPUs on the same scene which often boils down to AFR or SFR techniques which have a limit to how efficient they can be and/or hardware limitations like bus latency and throughput.

 

[flopper]

Takes time to write engine.
Learn how DX12 works and those writing code for Mantle will do better.
Most have had no issue adapting Mantle in a few months.

 

[Klunt Bumskrint]

Surely some devs will already be working on DX12 games. It's not like it's realised, gets delivered to their house and then they start.

 

[Silent_Scone]

Dx12 has native MGPU support so hopefully within time. Like with anything new it's letting developers have time to adopt their rendering tech

Takes time to write engine.
Learn how DX12 works and those writing code for Mantle will do better.
Most have had no issue adapting Mantle in a few months.

Lol do you even read what you type

 

[Mauller]

With the games rendering stack now controlling the gpu, explicit multi gpu will see scaling far closer to max theoretical. The problems with current multi gpu stem from the drivers not knowing what the game engine is doing. But this is no different to the problems that cause poor performance with current graphics apis. Requiring per game tweaks in drivers for best possible performance.

Multi gpu being hacked on top of dx is also another reason why some games never receive good multi gpu support. Some game engines just do not work well with either manufacturers drivers in this regard.

But it also depends on the rendering system used, AFR can see higher max fps but latency increases with every card added. Although total latency should be less than current methods.

SFR SuperTiling has the latency of a single card system, increased min and avg, but lower max compared to AFR. But I see it as the better system for scaling beyond two cards.

Multi gpu on Low Abstraction API's has spread show vastly superior performance in games like bf4 and Civ:BE using the AFR in Bf4 and SFR SuperTiling in Civ:BE.

 

[flopper]

Lol do you even read what you type

Not native english so get it wrong now and then think as me as Yoda do right find strength be happy not the way around.

SFR SuperTiling has the latency of a single card system, increased min and avg, but lower max compared to AFR. But I see it as the better system for scaling beyond two cards.

Multi gpu on Low Abstraction API's has spread show vastly superior performance in games like bf4 and Civ:BE using the AFR in Bf4 and SFR SuperTiling in Civ:BE.

SFR might be the main thing for multi gpu and we have 4 cards to scale in top titles. star citizen needs it

 

[Cooper]

DX12 is one of the reasons I want to hang on to my 970SLI setup, rather than selling them and getting 980TI SLI

Just curious to see if it unleashes more performance from them...if nothing else the new unified memory feature, effectively making them a single 7GB card is something I'm looking forward to.

 

[flopper]

http://blogs.msdn.com/b/directx/archive/2014/04/15/directx-12-developer-videos.aspx

https://channel9.msdn.com/Events/Build/2014/3-564

 

[Mauller]

DX12 is one of the reasons I want to hang on to my 970SLI setup, rather than selling them and getting 980TI SLI

Just curious to see if it unleashes more performance from them...if nothing else the new unified memory feature, effectively making them a single 7GB card is something I'm looking forward to.

People have a vast misunderstanding of the new unified memory system. They did not present it very well in their simplification.

What actually happens is that the memory on both cards is available. But the memory is still tied to an individual card. You gain full use of the memory of each card if the rendering system is using a method that processes different information on each card. so each GPU has their own memory workload.

TL;DR The memory of one card is not shared with another. Because there is no need to mirror RAM if AFR is not used, each card will have full use of its own memory.