Explicit multi GPU GTX1080/Fury X

[Acch3r0n (Blue)]

Hey, I have both a fury X and a GTX 1080 and I'm wondering how on earth you actually get both of these GPU's to work together in timespy. How do you enable Explicit multi gpu mode ?
Thanks

EDIT: Ok So I read that TimeSpy only supports mGpu which is a bit of a shame really and no actual AMD/Nvidia cross GPU action at all. OH well.

[UL_Jarnis]

Supports explicit multi-adapter, but only in linked node mode (LDA).

[Acch3r0n (Blue)]

Originally posted by FM_Jarnis:

Supports explicit multi-adapter, but only in linked node mode (LDA).

Are you guys going to be supporting explicit multi GPU with AMD/Nvidia GPUs in the same system ?

[UL_Jarnis]

Not with Time Spy. Doing so would require major rewrite of the engine. We did consider it when DX12 mGPU features were finalized by Microsoft and it was just too large change for extremely tiny user base ("those 5 people in the world with a modern AMD *and* NVIDIA GPU").

Doing it with dGPU + iGPU is not very practical - gains from it are fairly minor and compatibility headaches massive. It is very easy to have a "mGPU" setup that is suddenly slower than the dGPU alone...

Future tests ("in the far future"), it is possible. Partially this depends on how games adopt this possibility. 3DMark always aims to be a benchmark that does game-like loads and is as game-like as possible. If games do not adopt this (no, Ashes of Singlarity alone doesn't count... it is kinda the exception that proves the point), considering how complicated it is to program for it, it may not make sense.

[LBXZero]

By what I understand of DX12's multi-GPU mode, the explicit mode requires Crossfire/SLI to be disabled, or not linked in the driver. That way, DX12 takes control of each GPU individually instead of letting the driver manage it.

Basically, this "explicit multiadapter" support sounds like false advertising.

Edit, just read the technical PDF. Yep, Time Spy unofficially does NOT support explicit multi-adapter. It says so specifically in the technical manual, but it is worded to say it does support it. As said in the technical documentation, the LDA mode is implicit multi-adapter, letting the driver manage the multi-GPU handling.

[UL_Jarnis]

DX12 offers

MDA
LDA Implicit
LDA Explicit

Time Spy uses LDA Explicit.

http://www.bit-tech.net/hardware/graphics/2016/06/15/evga-geforce-gtx-1080-ftw-review/8

[LBXZero]

Why does LDA Explicit require Crossfire/SLI mode to be active in the driver to work if the engine is detecting and handling the 2 identical GPUs?

[UL_Jarnis]

That you have to ask AMD and NVIDIA driver developers. My educated guess is that it has something to do with what devices the driver reports as ready for rendering.

Time Spy asks "okay, what GPU is attached to the primary display?" and "Okay, what GPUs can render to this primary display?" and then starts dumping work on all of them via DX12 Explicit mGPU.

I would imagine that without Crossfire enabled, AMD driver responds to the second query "this one GPU here, have fun" and then only one would be used.

(without doing it like this, you could have the benchmark accidentally starting to run on the iGPU of the processor instead of the dGPU and then you would be making annoyed post about a really really bad score :D )

[LBXZero]

Basically, you are playing damage control, trying to prevent a fire from spreading.

You made a statement in the other thread on how MDA has no user base. That is entirely false. The average person would not have both an Nvidia GPU and an AMD GPU, but the average Intel CPU has an integrated GPU. For most consumers, they in fact do have a setup supporting an MDA configuration, being Intel GPU + AMD or Nvidia GPU. Alternatively, you have AMD APU + AMD or Nvidia GPU.

Next, you are poking fun at the benchmark accidentally selecting the iGPU instead of the dGPU. You obviously don't know what you are talking about. My laptop has that issue, but AMD and Nvidia have this thing called "software" that detects and selects which GPU is used based on power conditions. Really if there is a problem with 3DMark in choosing the right GPU, one the results will show it, and two, it is a design flaw in 3DMark as being a hardware benchmarking application. Being a benchmarking platform, 3DMark should be detecting the hardware and allowing options in custom configuration to select options.

Presently, it is obvious that Time Spy utilizes implicit LDA, not explicit LDA, as no option exists in Time Spy to control that, yet there is an option to disable Async Compute. If it is a software option, the option to enable or disable explicit LDA will be in the software.

[UL_Jarnis]

Problem with MDA is that if you do it like Ashes of Singularity, you get at best 2x the slowest GPU and everyone would be complaining why 3DMark isn't using all the resources of both GPUs. And to do it so that it would actually use all resources of two wildly different GPUs would require hideously complex code that would somehow "feel out" what the GPUs can do and then somehow load balance them, while responding to variances in performance and... uuuh, it would be very very complex piece of code and potentially a very fragile piece of code.

dGPU + iGPU the gains would be marginal and again the complexity would go through the roof.

I'm not saying a future 3DMark test can't ever support MDA, but Time Spy was designed for Linked-node. Partially because we wanted to ship it in 2016 and not 2017 and partially because mixed GPU (NV+AMD) setups basically do not exist outside press test labs.

Now if game developers suddenly decide that MDA is the best way to go and multiple games support NV+AMD mixed cards and everyone starts buying them so NV+AMD mixed setups are relevant for actual gaming, we'll definitely take notice.

(and to switch from Explicit LDA to Implicit LDA would require considerable changes to the engine. It is not a simple switch you could throw)

[UL_Jarnis]

As for selecting the GPU, the laptop switching software is indeed relevant, but in DX12 the application could actually do the switching itself. Problem is actually figuring out which GPU is which and what should be used. As far as I know (I'm not a programmer, so if this is suddenly a huge deal, I'll have to go ask for verification), 3DMark chooses to use the GPU that is rendering to the Primary Display in Windows (whatever you have set as primary display, attached to output of whatever GPU).

So if you had iGPU connected to a monitor and set as primary display, I guess Time Spy would run on the iGPU. On a desktop PC. Hence my "joke" about it somehow ending up using the wrong GPU.

On laptops all modern systems (exception: GTX 980 "notebook") the display panel is attached to the iGPU. dGPU rendering output is copied over to the iGPU and displayed by it. Unless you explicitly tell the driver to use dGPU for all rendering, Time Spy would default to the iGPU because that's who has the primary display attached. Current AMD and NVIDIA drivers do not have 3DMarkTimeSpy.exe yet pre-flagged as "use dGPU by default".

[LBXZero]

Originally posted by FM_Jarnis:

Now if game developers suddenly decide that MDA is the best way to go and multiple games support NV+AMD mixed cards and everyone starts buying them so NV+AMD mixed setups are relevant for actual gaming, we'll definitely take notice.

(and to switch from Explicit LDA to Implicit LDA would require considerable changes to the engine. It is not a simple switch you could throw)

According to several releases, developers are working on Intel+NV and Intel+AMD mixed setups.

According to Microsoft's details of DX12, explicit modes should not require Crossfire or SLI to be active.

[UL_Jarnis]

I too expected that Crossfire or SLI setting would not matter, but that is not how the hardware & drivers work. Sorry, not a driver developer, so don't know why.

As for supporting the mixed setups; We are predicting the future here, and did our prediction months ago. Our crystall ball says "made in [some random country I've never heard of]" so it may have been off in predicting that "MDA is too complex and the gains are too minor to support it commonly in games" and Time Spy being designed as a benchmark to reflect future DX12 gaming performance, decision was made to not use it.

I strongly doubt Time Spy is the last ever DX12 benchmark we do, so who knows what will happen in the future. Our guys are constantly following the world of GPUs and gathering input from all the GPU developers as well when doing future plans.

[LBXZero]

http://www.bit-tech.net/hardware/graphics/2016/06/15/evga-geforce-gtx-1080-ftw-review/8
According to this article regarding Nvidia's GTX 1080 on the MDA and 2 LDA modes, implicit specifically implies using SLI. Therefore, SLI and Crossfire should be disabled for explicit LDA to operate.

Time Spy's custom configuration does not offer the option to disable Explicit LDA. Therefore it is all official, Time Spy does not support explicit multiadapters. It cheats in saying explicit LDA support, unless there is an issue on adapter detection and registering when Time Spy loads.

So I guess you have a concern to report to your engineers. It is best for you to leave all questions unanswered regarding multi-GPU in Time Spy and wait for answers from your engineers and programmers.

[UL_Jarnis]

You are free to make your accusations, but I would suggest that you have presented no actual proof yet. Our development team stands behind what is stated in the technical guide.

"using SLI" is a bit imprecise term. "Using driver to control distribution of work across GPUs" would probably be more accurate. Time Spy does that in-engine and not by driver. Hence, LDA Explicit.

Now I honestly don't know if the issue is just that Time Spy can't see the additional GPU(s) if SLI/Crossfire is disabled and that is why it cannot distribute work to more than one GPU. That is possible. I guess I could try ask from the developers if they happen to know.

[LBXZero]

http://forums.ashesofthesingularity.com/475804

For explicit modes to work, Stardock recommends Crossfire and SLI disabled for best performance. The problem is obviously only on Futuremark's side. You may have been given false information to cover up a marketing error.