During AMD's presentation at Computex 2016, the company showed no fear as it compared the new Polaris-based Radeon RX 480 GPU to Nvidia's GeForce GTX 1080. The $700 Pascal-based powerhouse was tested in Ashes of the Singularity versus a pair of the Polaris GPUs and came away a little embarrassed in this scenario. AMD presented a slide that claimed the GTX 1080 only managed 58.7 (we presume average) FPS, while the Radeons pushed out 62.5 FPS using the same settings.
Sharp-eyed viewers, including a few of you here on TR, noticed some oddities in the numbers and in the video stream. Units and world geometry weren't in quite the same places, and one of the renders appeared to be using different quality settings from the other. AMD also claimed 98.7% GPU utilization for the GTX 1080, while the RX 480 cards were supposedly only 51% loaded. What was going on?
Well, yesterday, Robert Hallock, AMD's Global Head of Technical Marketing, made a post in the official AMD sub-Reddit attempting to explain what was causing the confusion. He stated that both tests were run on the same Core i7-5930K-equipped machine, differing only in GPU configuration. He also confirmed that the tests were run with the same settings: "Crazy" quality in 1080p with 8x MSAA and v-sync off.
For starters, Ashes of the Singularity uses a random seed to procedurally generate certain game assets, leading to slight variation in the way the benchmark looks. More interestingly, though, Robert says the GTX 1080 is "incorrectly executing" some of the shader code in the benchmark, which leads to less snow pile-up on the terrain. In his words, "the GTX 1080 is doing less work to render AOTS than it otherwise would if the shader were being run properly." According to Robert, the RX 480 is correctly executing the terrain shaders, leading to more procedurally generated snow—and worse performance.
Robert did not go to great lengths to explain the GPU usage metrics. He did however say that the RX 480 was CPU-limited in the "single" or "normal" batch test, and that that was the reason for the low 51% GPU usage number given in the presentation. GPU usage did climb to 71.9% in the "medium batch" test, and again to 92.3% in the "heavy" batch test. Robert stated that the final multi-GPU scaling in these tests was 183% versus a single GPU.
Ashes of the Singularity runs on DirectX 12 and uses Explicit Multi-Adapter mode for multi-GPU processing, so the GPUs were not running in Crossfire mode. As such, these results are not actually indicative of Crossfire scaling on the RX 480. Still, this performance level is impressive for a pair of $200 graphics cards. We can't wait to see what just one can do.