Page 2 : The Odd Results Explains, Summary Of Key Points, What Does This Mean?
The Odd Radeon RX 480 AOTS Results Explained…
Robert Hallock was then asked about the odd 51% GPU utilisation result. He clarified that it was based on the single-batch (normal batch) data set, which was CPU-limited.
The single-batch or normal batch data set is about 60 seconds of scenes with lower (typically under 10,000) draw call counts. The medium batch data set consists of another 60 seconds of scenes with typical draw call counts between 10,000 and 20,000. The final heavy batch data set tests scenes with draw call counts in excess of 20,000.
To clarify this, the scaling from 1->2 GPUs in the dual RX 480 test we assembled is 1.83x. The OP was looking only at the lowest draw call rates when asking about the 51%. The single batch GPU utilization is 51% (CPU-bound), medium is 71.9% utilization (less CPU-bound) and heavy batch utilization is 92.3% (not CPU-bound). All together for the entire test, there is 1.83X the performance of a single GPU in what users saw on YouTube. The mGPU subsystem of AOTS is very robust.
Let Us Summarise The Points…
Based on his explanation, we can derive the following conclusions :
- AMD claims that the GeForce GTX 1080 is performing a few percent better than it really should because it was improperly rendering the terrain in AOTS.
- Two Radeon RX 480 graphics cards will deliver 83% better performance than a single Radeon RX 480 graphics card.
- The benchmark scores are the average for the entire AOTS benchmark, which consists of the Normal Batch, Medium Batch and Heavy Batch data sets.
- The 51% GPU utilisation was cherry-picked from the Normal Batch / Single Batch results.
- The frame rate for Normal Batch / Single Batch data set was not revealed, but it would have been much, much lower than the 62.5 fps average frame rate.
- The average GPU utilisation for the two AMD Radeon RX 480 cards was 71.7%, while the average GPU utilisation for the NVIDIA GeForce GTX 1080 [Amazon] was 98.4%.
In short, the controversy was created by AMD cherry-picking the GPU utilisation of the most CPU-limited portion of the AOTS benchmark, while comparing the average frame rate achieved in the entire AOTS benchmark.[adrotate banner=”5″]
What Does This Mean?
Despite what Raja Koduri said in the video clip, there isn’t “much, much more headroom” for developers to squeeze out of the Radeon RX 480. That 49% headroom was only seen because the test was CPU-limited.
The two Radeon RX 480 cards were literally “chilling” 49% of the time, because they were waiting for the CPU to finish processing the AI and physics, before rendering a new frame. In fact, they most likely delivered a much lower frame rate as a result of being CPU-limited.
AMD inaccurately compared the average frame rate of the entire Ashes of the Singularity benchmark with the GPU utilisation of the Single Batch / Normal Batch portion of the benchmark (which is CPU-limited). Here is our rough correction of their comparison, based on the average GPU utilisation for the entire benchmark :
The adjusted results are still impressive, because it shows that there is still some headroom. Just not as much as was presented. We have no idea why they needed to cherry-pick that result, since the Radeon RX 480 will still impress gamers from the performance/cost point of view.
Based on our extrapolation of their comments, we can guesstimate that a single AMD Radeon RX 480 graphics card will deliver an average frame rate of 34 fps in the Ashes of the Singularity benchmark. That means the NVIDIA GeForce GTX 1080 Founders Edition [Amazon] is about 72.6% faster than the Radeon RX 480.
That sounds about right, and is hardly surprising since the GeForce GTX 1080 Founders Edition [Amazon] costs US$699 while the Radeon RX 480 (8 GB) will only cost US$229. There is nothing for AMD to be ashamed of achieving “only 58%” of the GeForce GTX 1080 Founders Edition’s performance. After all, it costs less than 1/3 of the price!