Thank you for the latest build. Here is my testing with this build.
First problem CUDA error rendering with using this blender 2.79 benchmark with my 1x rtx2070. Here is the link The new Cycles CPU 2.79 Benchmark.
CUDA error: Illegal address in mem_alloc_result, line 850
Error: CUDA error: Illegal address in cuCtxSynchronize(), line 1686
I did some benchmarking with your latest build.
BMW scene - 1x rtx2070
Release_E_Cycles_win_RTX_20190103 (auto tiles): 46.66
your latest (auto tiles): 47.36.
your latest (256x256): 57.06.
Classroom scene no compositor and no layers - 2x rtx2070
Release_E_Cycles_win_RTX_20190103 (auto tiles): 1:25.73
your latest (auto tiles): 1:29.87.
your latest (256x256): 1:29.57.
The latest build is little slower compare to previous one.
Another issue is in some scenes like in the BMW scene the auto tiles is noticeably faster than (256x256) tile. I prefer using the 256x256 because the denoiser is a lot faster. Could we improve performance for the 256x256 tile?
In the builds would it be possible to expose the scrambling distance parameter so I can do some testing with it? Now that you have expose dithered sobol.