I just spent the last hour testing K-Cycles and have used E-Cycles for all of 2020. I no longer own E-Cycles so I couldn’t compare them directly.
The Viewport render improvement is AWESOME by the way and almost worth the full price by itself.
Vanilla Cycles 2.92 vs K-Cycles 2.92
1920x1080 @ 2000 samples
RTX 3090 x 2
Default High Quality Preset
Auto-Tile Size On
K-Cycles(Left) - Vanilla (Right)
Viewport Test @ 256 samples
Vanilla 33 secs
K-Cycles 19 secs
Viewport Test @ 256 samples
Vanilla 29 sec
K-Cycles 16 secs
Viewport Test @ 256 samples
Vanilla 34 secs
K-Cycles 28 secs
NOTE: This last test was invalid. Accidentally left NVlink mode on for the K-Cycles render. Second time around K-Cycles was almost 30 seconds faster.
Image quality is nearly identical to Vanilla but not always. In two of these renders I noticed a very slight difference which I then confirmed in scopes.
(These tests were for myself as I had no intention of posting them so I apologize ahead of time for not having them all perfectly side-by-side).
Possible Bugs: the viewport denoiser is bugged on my end. Denoising doesn’t happen until sampling is completed no matter what Start Sample is set to.
Task Manager Performance tab set to Cuda and Compute-1 show both GPU’s rendering at 100%. As soon as one tile finishes, one of the GPU’s drops to 0%. It would seem that one card becomes completely inactive while waiting for the 2nd tile to finish rendering. Visual bug?
Additionally, you won’t get the full performance boost without using Auto-Tile Size. This is a bit annoying as the final render is done in two massive tiles so you don’t see the results till the end. Not a biggie but worth mentioning.
Conclusion: Buy it. It’s more or less the same performance of E-Cycles with an even better viewport. None of the features of E-Cycles like Light Groups/Persistent Data. Unless the features of E-Cycles are important to you, for a fraction of the cost, this is a clear winner.
Edit: Now with Light Groups & Light Linking, K-Cycles is the clear winner for me.