The First DXR Benchmark Has Arrived
Microsofts DXR launched roughly a year ago and Nvidias Turing based GPUs about four months ago. While both are relatively young, its hard not to be anxious about testing out the new tech that both have brought. Up until recently on the ray tracing front we have had only a few samples (that didn't look impressive) from MS and finally Battlefield V. With the release of Port Royal by UL Benchmarks (formerly Futuremark) we now have another test in the very limited arsenal that we can use to see just what this new tech and these GPUs are capable of.
Instead of running just the base test, we decided to go a bit deeper and run all custom options as well. There was no crazy insight found by running these tests, but more a show of what each setting cost in the overall scheme of things. I can point out that Anisotropic filtering had absolutely no impact on performance on the 2080 Ti, turning off all reflections had a very interesting look to it (everything that was reflective was no black as opposed to non reflective). Turning off transparencies darkened up the scene quite a bit as well.
So, while we wait for the DLSS patch to come out (which we were told would be 'soon'), here are the tests for each of the settings.
First up we have reflection mode. This setting allows for ray traced reflections, traditional (SSR / Cubemap) reflections, and then finally no reflections. Disabling ray traced reflections nets a 15% increase in performance, while disabling all reflections (and I believe this disables all ray tracing in the engine as well) nets a 215% increase in performance over ray traced. Though lets not kid ourselves here, this is not representative of a real world gaming workload as its not visually appealing at all and does not look like a game at all.
Next up is reflection quality. By default reflection quality is set to medium. Increasing the quality to high results in a 12% loss in performance. Going in the opposite direction, medium to low results in a 11.3% increase in performance. Turing quality to off simply resulted in a crash.
Here you can see the performance with async disabled. With traditional reflections, the score with traditional reflections and async on (sorry for leaving it out of the chart) was 9343. This is a 4% reduction in performance. With Reflections enabled, you see a 3% reduction disabling Async compute. It seems like this is a very light implementation and its possible that ray tracing simply doesn't leave room for very much asynchronous operation as its very shading heavy.
The other ray traced effect in this benchmark is shadows. This in my opinion added the least visually to the scene and with what we've seen, I find this to be the least valuable RT effect of the 3 (Reflections, Global Illumination, Shadows). That said, disabling shadows resulted in a gain of 4% gain in performance, while disabling reflections and shadows resulted in an increase of 19% over stock. So overall it looks like the ray tracing technology in this bench altogether results in a 20%~ hit in performance.
Another interesting and telling setting was reflection sampling. This effectively is how many rays you are casting per pixel (these are only calculated further upon hit detection). This setting increases the quality of reflections pretty substantially and showed just how much we have to gain by increasing GPU performance to allow more samples. On to the numbers. From stock to two samples, we get a reduction of 14%, going to 3 samples we lose an additional 12% of performance. Finally, jumping up to four samples and we have an additional 11% decrease in performance. As we can see, the reduction drop scales almost linearly so we know that the performance increase required scales as such.
Here we have all the basic effects disabled in turn measuring our base rasterized (traditional shading) performance. I won't list the performance differences here as the focus was primarily on ray tracing performance, but I will say that disabling transparents was an interesting effect.
Testing this was interesting in the fact that TAA sampling while having a noticeable affect on image quality, had no real impact on performance when changing between on and off. Even turning it off the performance gain was negligible and only netted a 30 point increase. All of this said, tensor cores are not used at all for denoising and is processed simply using the TAA pass, much like Battlefield V
Well, that was it for this test. This was an interesting test to do, but not a very enlightening one. We can conclude that with ray tracing disabled we only get a 20% increase in performance here, which is roughly 10% better than what we have in BFV with it turned on. That said, Battlefield V has a much more dynamic environment with much more going on in it. I expect that in environments like that, we will continue to see performance hits that line up with theirs. Hope that this testing gave some insight to some of you. Until next time, PC Better.