Frame Rating: GTX 970 Memory Issues Tested in SLI
Battlefield 4 Results
At the end of my first Frame Rating evaluation of the GTX 970 after the discovery of the memory architecture issue, I proposed the idea that SLI testing would need to be done to come to a more concrete conclusion on the entire debate. It seems that our readers and the community at large agreed with us in this instance, repeatedly asking for those results in the comments of the story. After spending the better part of a full day running and re-running SLI results on a pair of GeForce GTX 970 and GTX 980 cards, we have the answers you're looking for.
Today's story is going to be short on details and long on data, so if you want the full back story on what is going on why we are taking a specific look at the GTX 970 in this capacity, read here:
- Part 1: NVIDIA issues initial statement
- Part 2: Full GTX 970 memory architecture disclosed
- Part 3: Frame Rating: GTX 970 vs GTX 980
- Part 4: Frame Rating: GTX 970 SLI vs GTX 980 SLI (what you are reading now)
Okay, are we good now? Let's dive into the first set of results in Battlefield 4.
Battlefield 4 Results
Just as I did with the first GTX 970 performance testing article, I tested Battlefield 4 at 3840x2160 (4K) and utilized the game's ability to linearly scale resolution to help me increase GPU memory allocation. In the game settings you can change that scaling option by a percentage: I went from 110% to 150% in 10% increments, increasing the load on the GPU with each step.
Memory allocation between the two SLI configurations was similar, but not as perfectly aligned with each other as we saw with our single GPU testing.
In a couple of cases, at 120% and 130% scaling, the GTX 970 cards in SLI are actually each using more memory than the GTX 980 cards. That difference is only ~100MB but that delta was not present at all in the single GPU testing.
Our performance data is being broken up into two sets: the GTX 980s in SLI running at all five of our scaling settings and, separately, the GTX 970s in SLI running on the same five scaling settings. Plotting 10 sets of data on a single graph proved to be a a bit too crowded, so we'll show the graphs successively to help you compare them more easily.
Unlike our first sets of results the SLI numbers are ALMOST in a playable state, making them much more real-world than before. The first thing I noticed when compiling this data was that the GTX 980 cards in SLI actually had a couple of more downward spikes in frame rate at 150% scaling than the GTX 970s did. I did confirm this was a regular pattern by re-running tests on both sets of hardware about six times, and the bright green line you see in the first graph above is actually one of the better results for the 980s.
It appears though that moving from 110% to 150% scaling results in the expected frame rate decreases in both configurations.
Average frame rates are where we expect them to be: the GTX 980 SLI is faster than the GTX 970 SLI by fairly regular margins.
|GTX 980||GTX 970||% Difference|
|1.10x Scaling||47.3 FPS||41.0 FPS||-15%|
|1.20x Scaling||41.1 FPS||35.8 FPS||-15%|
|1.30x Scaling||35.4 FPS||31.2 FPS||-13%|
|1.40x Scaling||31.0 FPS||27.7 FPS||-12%|
|1.50x Scaling||27.7 FPS||24.6 FPS||-13%|
The percentage differences here are actually more reliable than the single GPU results, which is a bit of a surprise to us. The GTX 970s in SLI range from 12-15% slower than the GTX 980s in SLI, but as we know from our years of GPU evaluation, that isn't the whole story.
Click to Enlarge
It should be painfully obvious that at 150% scaling the GTX 970s in SLI have a significant amount of frame time variance that is mostly limited with the GTX 980s in SLI. Even at 140% scaling, looking at the thinner gray line, you can see differences in the behavior of the frame time lines in the two graphs (you can click to enlarge them for a closer view).
Before you start eyeballing these graphs, make sure you take note of the slightly different y-axis on the left hand side between them - that's important. Focusing on the highest 10% of frame times in our testing, the GTX 970s clearly exhibit more issues than the GTX 980s. The flagship cards only see about 5ms of variance at the 90th percentile and crosses the 20ms mark somewhere in the 96-97th percentile. The GTX 970s in SLI though reach those marks much sooner - at the 90th percentile we see as much as 18ms frame variance and at the 96th percentile that reaches as high as ~40ms.