To make some memory, when we saw the AFR SLI and CrossFire We understood that the way to work was to distribute each GPU frame, so that one graphics card had two and the other one was complicated. Is this best and most effective? What does SFR contribute to its competitor and how does it work?
Distributing a Split Framework: a process that does not work well, but is used
If AFR is able to assign each frame to a graphics card and divide the work in sequence, the SFR is probably the opposite. Instead of separating each frame of the GPU in question, SFR divides the screen into two or more sections (depending on the number of graphics cards in the system) directly, as if we're talking about a retro multiplayer game.
One GPU will represent and render part of the screen and another GPU the opposite. Obviously, this idea sounds great, but we all know that, in play, often one part of the stage is more loaded than the other, so the GPU will have more work and bring the frames seven more miles later than his sister's.
SFR by its nature, it has a suitable dynamic loading system that works on a compatible driver, so that, if one GPU is overloaded, another GPU will help. This generates a huge workload, since both GPUs have to constantly rotate the image position to try to make the task much bigger.
The problem is that many times such a task can be repeated with a very small change of location, so the maximum communication delay is large and you will need a direct bus such as SLI or NVLink
Why is SFR still used when performance is not so high?
With the mechanics of the industry basically. Most developers work very easily when using SFR and AFR either Hybrid SLI either SLIAA. From the perspective of the programmer and engineer used, it is very easy to work knowing that the load distribution planning will be based on the partitioning of the screen and not all of the components created.
In fact there are companies and engines that work so well with SFR that they are even popular with AFR because of the number of hours and efficiency developed by the engine over time.
Obviously, these examples are few, but surprisingly when great accomplishments are achieved, the final outrage is usually low, even though performance is not as good as it is with AFR.
So, we're talking about some very specific and very specific cases, something SFR hasn't quite mentioned as providing technology for systems with more than one GPU. In fact, AMD for example uses SFR more in the DX12 that in DX11, when it was reliably deployed in AFR because it didn't require an independent line, something to gain thanks to asynchronous computing, something that was discussed a lot when it came out with version 12 of the Microsoft API.