I decided to update my findings on the performance of the three variations of GCN that have identical top-level CU configurations, Tahiti XT, Tonga XT, and the Polaris 20 (PRO) chip. Here I tested The Witcher 3.
The scene is the introduction with Geralt in the bathtub. Because it's convenient as it's right at the beginning and rendered in real-time. And Geralt's hair is simulated with HairWorks tessellation. Benchmarking begins when the scene first renders the fire to when Geralt throws the creature thing out of the bath and it hits the floor and fizzles out.
I did this test twice, once at maximum settings and then again at 720p resolution, and low post-processing. That second one was to try and shift the bottleneck towards the primitive-throughput of hte GPU instead of shader/bandwidth. But the first test should also let us see how well these GPUs do with the exact same core clock and memory bandwidth.
First is 1920x1080 resolution.
Total increase in performance in this test from GCN1 to GCN4, with the same number of CU, ROPs and core clock rate and memory bandwidth is 24.9%. Not too bad.
Let's take a look at the 720p results with the emphasis on geometry performance over shading.
A 32.6% increase in generational performance between 1 and 4 iterations of GCN. So the performance gap did widen as we try and emphasis triangle throughput.
So that concludes this little micro-test.