The NVIDIA RTX will complete a month from its presentation and the truth is that few data can be extracted from its performance under Ray Tracing. Let’s not talk about comparing with other architectures and much less making a comparison with an AMD graphics card, until today. Today, the real difference between the NVIDIA computing units versus the previous architectures under Ray Tracing can be accounted for.
Everything points to an unpublished driver from NVIDIA that allows the Microsoft DXR API to calculate ray tracing without the hardware acceleration that Turing has and that the guys at Hardwareluxx have been able to catch it.
As is well known, Pascal (and the architectures prior to it) does not have exclusive ray tracing units ( RT Cores ) as if he has Turing and therefore the ray tracing calculation has to do it “by brute force”.
This means that they have to use their shaders for those calculations at the same time they render, with the consequent loss of performance against the RT Cores de Turing.
To this is added the fact that they also do not have AI units ( Tensor Cores ) that are used in Turing for the new DLSS.
So what performance can an architecture as powerful as Pascal have without those two processing units versus Turing?
The data we are going to show is based on an impartial API like DXR that runs on Windows 10 through Direct Compute, and the unit of measurement used is frames per second.
The software chosen for comparison has been the famous Star Wars render under Unreal Engine 4 that allows RT and DLSS.
On this occasion an intermediate resolution has been chosen, such as 2K, to which Ray Tracing and DLSS have added, revealing that Pascal can not render the scene at 10 FPS, while the two RTX are able to move it more or less well, giving results of 43.6 FPS and 55.4 FPS respectively.
If we raise the resolution and leave Ray Tracing alone, the data logically worsens but the percentage differences remain. The RTX 2080 Ti gets 34.6 FPS while the RTX 2080 achieves 28.4 FPS. In contrast, Pascal cards barely exceed 5 FPS, in the case of the GTX 1080 Ti just 5.5 FPS.
This means that Turing with its RT cores and Tensor cores is almost 7 times faster than Pascal under Ray Tracing and DLSS (GTX 1080 Ti vs RTX 2080 Ti), data that is maintained if we only leave Ray Tracing enabled. If we compare cards with a very similar performance, such as the GTX 1080 Ti and the RTX 2080, the difference is just over 5 times higher, which is not bad.
Right in the middle is the Titan V, which is a card that only has Tensor Cores but not RT Cores, although it is much more powerful than the Pascal GPU. This only evidences one thing, you need unique processing units for “real time” ray tracing.
The Microsoft API manages to do a good job since the differences so stipulate, but we do not know (at the moment) what performance the cards shown here under the NVIDIA API can have. In any case, NVIDIA has been fully successful with the Pascal architecture as far as RT is concerned.