Testing multiplayer games involves a lot of finicky variability, so I stuck with Battlefield 4's single-player campaign. Benchmarking was done at the start of the Singapore beach landing, which features cataclysmic environmental effects as well as explosions, gunfire, and... well, see for yourself:
I tested at 1366x768 using the "Low" detail preset. The A8-7600 did fare reasonably well at 1080p with the "Medium" preset in Geoff's testing, but that was in a different section of the game. The Singapore beach landing takes place outdoors, and it's quite demanding from a graphical standpoint. I used this same stretch of the game to test the Radeon R9 270 last November.
These results look an awful lot like those from Scott's first look at Mantle. On Kaveri, just as on Scott's R9 290X-powered system, the Mantle renderer yields a somewhat flatter plot that sits lower along the Y axis, indicating faster, more consistent performance. At the same time, the Mantle renderer produces very high latency spikes on occasion. The graph above is clipped at 160 ms, but the spikes you see reach just over 200 ms. At least one such spike occurred in each of the five test runs conducted.
Let's what we can learn from slicing and dicing this data into some different graphs.
The average FPS and 99th-percentile graphs highlight one of my observations from the frame-by-frame plot: the Mantle renderer does yield better performance on average than the Direct3D one. The difference isn't very big on Kaveri, but it's there. We'll talk some more about subjective impressions in a minute.
The latency curve and "badness" thresholds show, in detail, the flip side of the Mantle renderer's performance increase. On the A8-7600, occasional latency spikes cause the Mantle renderer to spend more time above 50 ms than the Direct3D one. In spite of that, Mantle yields enough of a performance and smoothness increase that it spends less time over 33.3 ms. That's a double-edged sword if I ever saw one.
(We also have a graph showing differences above 16.7 ms, but in this case, the data are more academic than anything. The A8-7600 spent more than half of its 60-second run above the 16.7-ms threshold regardless of the renderer used.)
All right. Now that we've looked at the cold, hard numbers, let's talk about the subjective, seat-of-the-pants impact of the Mantle renderer on Kaveri—and discuss whether or not you should use it just yet.