Go Back   OC3D Forums > [OC3D] General Forums > OC3D News
Thread Tools Display Modes
Old 12-02-19, 11:30 AM
tgrech tgrech is offline
OC3D Crew
Join Date: Jun 2013
Location: UK
Posts: 603
DX12 & Vulkan bypass the drivers(And therefore the traditional mechanism of SLI/CFX) but can view a multi-GPU set up in "Linked node" mode which allows some degree of compatibility with traditional AFR implementations. In theory you can use two GPUs however you'd like in a DX12/Vk setup, which in theory opens the door to all sorts of exotic mixes and weird techniques. In practice, 90% of these will be incredibly unoptimal, hard to balance and take far more work to get right than they're worth. NVidia, AMD & MS have put some work into developing multi-GPU libraries(A collection of premade software for developers to use to speed things up) for DX12/Vk that re-implement more mature techniques like AFR but balancing these workloads in the game code for every type of GPU has turned out to just turn an already kinda messy solution into a much less workable one. While in theory low-level APIs have given us an oppurtunity to replace old AFR techniques with new possibly better ones, most effort is now just back to trying to get those old ones working in a basic sense first, and it's looking like both hardware vendors and software developers are just not seeing the benefits for it in gaming to chase up creating such complex infrastructure, still preferring to revert to driver-managed AFR(I guess just think of how long it took hardware vendors to get reasonable multi-GPU setups and then consider game developers are smaller teams trying to create software for far more hardware and you can see why it quickly becomes a flawed approach). This article discusses some of those techniques https://developer.nvidia.com/explici...ectx-12-part-2

But, part of the lack of interest is because one of the hottest topics in microprocessor design and research at the moment is multi-die single processors, like Epyc2, that use a collection of different building blocks that can be put together tightly enough that they can process a single workload coherently with only optimisation work required for the non-uniform memory access considerations. The bandwidth of links like Infinity Fabric isn't nearly there yet for carrying this tech over to GPUs, but every major tech company has announced this is without doubt the direction we're heading in over the next couple years and there's a lot of new techniques in the area of linking these dies together.

Electrical engineer with a passion for rambling.

Check out my latest piece of commissioned work: http://ehenleyshutters.com/novo-in-black/
[It's a company page not run by me, I just did the electronics & software design work]
Reply With Quote

Thread Tools
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump

All times are GMT. The time now is 06:33 AM.
Powered by vBulletin® Version 3.8.7
Copyright ©2000 - 2019, vBulletin Solutions, Inc.