HHbomb wrote:RTX cards are really fantastic, but what about rendered image quality ? Same as before ?
Yes - as mentioned in the FAQ quality is the same. The worst case scenario is that for scenes that don't benefit from RT Cores (volumes and non-mesh heavy scenes, low complexity mesh scenes), they render at least the same OB as before (e.g. RTX off). The scene in this RTX benchmark is the first one we encountered (basically by accident during testing) that always hits around 2-3x the speed, most were half that.
To be clear, and to make sure we set expectations correctly, RTX boost IS NOT going to get you a 2.5x+ increase in path tracing speed (or 5x+ the info kernel speed) for every scene. Today, it is consistently delivering a little less than half this boost for many scenes we've tested (e.g. 1.2x+ in PT and 2x in Info kernel). On the plus side, we never got anywhere near a 3x PT/DL speed boost in Octane just a few months ago, so with further tuning we hope to see if we can get more consistent high speed boosts across more and more scenes...
The purpose of this benchmark today is to provide a current WIP snapshot of the scene dependent min/max range for rendering speed on RTX GPUs. The min/max range starts with normal OB (i.e. RTX off= approx OB3/OB4 speed) as the minimum OB speed you should count on for all scenes, and 'OBX' as our maximum possible OB speed (with RTX on) as measured in best case scenarios like in this benchmark scene, which can fully benefit from RT Core HW.
Our work going forward, beyond engine tuning, is to provide more RT Core optimization tools and features that make it easier to leverage full 'OBX' speed. One example, the new 2019.1 vertex displacement system can fully use RT Cores, but old (pre-2019.1) displacement nodes do not (and will be relatively much slower on RTX). Down the line, we are looking at a scene optimizer (i.e. converting old displacement nodes to the new system mentioned before) or even a completely new kernel that could be added alongside current PT/DF/PMC kernels, but with minor (possibly automated) workflow changes is designed for 'maximum OBX' speed on RT Cores (more on that at GTC).