At my working place we’ve just got the Titan RTX and the IT dep is kindly suggesting to use it in windows 10, but after a couple of tests I’ve got slightly disappointed. More precisely I ran the pytorch-gpu-benchmark (https://github.com/ryujaehun/pytorch-gpu-benchmark) and all the numbers are twice the ones shown for TitanV…moreover I read the TCC vs WDDM thread on stackoverflow (https://stackoverflow.com/questions/17389179/is-there-a-way-to-turn-off-windows-display-driver-model-wddm-in-windows-7-8). I there any real issue on using windows instead of linuf or am I missing something here?
Certainly the WDDM issue on Windows is real, and for serious compute work I would not want to use a WDDM GPU on windows. It should be possible to put Titan RTX into TCC mode on windows.
The deep learning software stacks are generally tuned on linux.
Generally speaking, running a GPU with the TCC driver on Windows typically results in equal, or even minimally higher, performance with running that same GPU on Linux, with all other system components staying the same. The WDDM overhead issues have gotten worse with WDDM 2.x which is used since Windows 10.
From extensive experience developing software both on Windows and Linux, I would claim that a Linux environment offers higher productivity, reducing development time by as much as 25%.
I have got the some benchmark results both in Windows and Ubuntu: much lower than Titan V.
It could be a firmware/driver problem?
I would expect a FP64 benchmark to be much lower on Titan RTX than Titan V.