Click through the buttons above to see frame-by-frame results from a single test run for each of the graphics cards. You can see how there are occasional spikes on each of the cards. They tend to happen at the very beginning of each test run and a couple of times later when I'm exploding dudes with dynamite arrows.
You can see from the raw plots that the 290X looks good, with more frames produced and generally lower frame rendering times than anything else we tested. Every card encounters a few slowdowns, and the spikes on the 290X aren't anything exceptional.
The traditional FPS average and our frame-latency-focused companion, the 99th percentile frame rendering time, pretty much agree here. That's a good indication that none of the graphics cards are encountering any weird issues. When they don't agree, as sometimes happens, bad things are afoot. What they agree on is simple enough: the 290X is the fastest graphics card in this test. The uber fan mode doesn't seem to make much difference here.
We can get a broader sense of the frame time distribution by looking at the tail end of the curve. In this case, both brands of GPUs, faster and slower models, all suffer from a small number of high-latency frames in the last ~2% of frames rendered. I suspect the performance problem here is at the CPU or system level, not in the graphics cards themselves, since it's fairly consistent.
Our "badness" index concentrates on those frames that take a long time to produce. For the first two thresholds of 50 and 33 ms, the results are pretty similar among the newer GPUs, which again suggests a CPU bottleneck or the like. However, for slinging out frames 60 times per second, once every 16.7 milliseconds, the R9 290X is easily the best choice.