SLI Scaling In Synthetic Benchmarks
We've all seen plenty of SLI scaling benchmarks, where the analysis focused on some piece of test data, such as "SLI scaling in game x at preset y is z%". Unfortunately, reality is more complex than that, and almost all of those single-number tests don't tell you the whole story. Without some sort of context on what is limiting scaling, it's almost impossible to truly understand the technology's potential.
Together, let's explore SLI scaling in more detail starting with simple synthetic benchmarks. Later, we'll move on to real-world games.
Unigine Valley's Extreme HD preset is run at 1080p (with 8x MSAA). As we'll see from testing other settings, above about 150 FPS, the system's bottleneck shifts from the GPUs to the CPU. Hence the apparent scaling of "only" 73% in this scenario.
What's really happening is that scaling for most scenes is actually close to 100%, while some of the highest-FPS sequences are CPU-limited with SLI active. This leads to a lower average for that benchmark.
Notice that, at about 12:21 PM in the data, very high FPS (above 150) are coupled with decreased GPU core utilization. That's our evidence that the bottleneck is shifting (in this case, to the CPU). I also want to point out that, despite the 980's lower memory bandwidth, the video memory controller is never a bottleneck in this test; it hovers at 80% utilization. Third, look at how low the PCIe link utilization is, even though it's only eight lanes wide. We're reporting about 10% there, though that metric is considered inaccurate by Nvidia and not used internally.
Now for more demanding resolutions...
In order to properly assess SLI scaling, we want to give the GPUs a more taxing workload. How about the Custom Graphics Test 2 of 3DMark Fire Strike at 1440p with 16xAF, 8x MSAA and all of the eye candy turned all the way up? In this scenario, we see closer-to-ideal 88% scaling.
It's time to really punish those GPUs by turning the resolution up to 2160p (also known as 4K) with 16x AF, no AA and all of the detail settings maximized. Scaling is an almost theoretically-ideal 96%, though, in this extreme test, we don't observe acceptable framerates, even with two GeForce GTX 980s in SLI.
Note that custom runs of Fire Strike do not produce a numerical "points" score. They should be interpreted as relative to each other.