Performance Improvements in last several Driver releases

Accelerated Computing CUDA CUDA on Windows Subsystem for Linux
#1

I see claims of performance improvements in latest driver release notes - yet I for one am not seeing any significant performance increases of late - may be 5-10% at most since last several upgrades. I have been running “ai-benchmark” suite - I see score of 15k for RTX 0270 on WSL2 vs 25k on Linux. Similarly In my real-world pattern-matching application - WSL2 is much slower than Linux on same hardware (RTX 5000).

I can accept 5-10% slowness but 40% is just too much. What is ultimate goal of Nvidia WSL2 Driver effort? To be competitive w/ Linux? Any hope if we are going to get there soon?

#2

Thank you for the feedback. We are glad to hear that you are interested in measuring WSL Performance improvements.

To obtain the latest performance results, please be on latest WIP OS from dev channel : 21313 and ensure the WSL Linux Kernel is 5.4.91

More detail guide how to install and update WSL is documented by Microsoft can be found here:

The kernel update is also available as MSI here:

Please note that this is a recent WIP OS, we have not fully validated against it yet.

You will soon here from us on our performance goals and the work we have done in that front.

Thank you for reaching out!

#3

Thank you for your reply! I installed the latest Kernel and indeed there is an impressive jump of 33% in performance (my 2070 RTX now reports 20k in ai-benchmark vs 15k before. WSL is now just 20% behind Windows or Linux (24-25k range for this card). This is very impressive. Please keep us posted. It will be nice if the gap got to within 5-10%.

#4

That’s great to learn that you were able to reproduce some of our performance improvements.

Could you provide more details on the particular benchmark and whether you have Hardware Scheduling enabled as well ?

Performance improvements are highly workload dependent. Could you try other performance benchmarks of longer workload as well?

Thank you and we are looking forward to hearing from you

#5

Benchmark I am running is called Tensorflow Ai-benchmark (its a quick 10 min benchmark see: https://ai-benchmark.com/ for more) Its not the most rigorous benchmark but it its quick and good enough I would say.

I am not sure if Hardware Scheduling was on before but I ran same benchmark making sure Hardware Scheduling was on and this time it WSL2 came within 4% of Windows.

WDDM Windows Ai-Benchmark Score for Super 2070: 22433
WDDM WSL2 Ubuntu Ai-Benchmark Score for Super 2070: 21548 (4% less)

TCC mode for Quadro cards like my RTX 5000 gets you another 5-7% but WSL driver only virtualizes in WDDM mode . SO its fair to say its likely no more than 10% slower vs WIndows even when Windows is run TCC mode and WSL in WDDM mode.

I am now going to run same benchmark on Linux (same machine same NVME drive - in dual boot mode)

I also have some longer heavy duty OpenCV template matching jobs (that runs 2-4 hrs on multiple GPUs - quite intensive) and some faster rcnn models that run much longer. So more results coming.

#6

Score for Ubuntu 20/Linux on same machine: 24012
So WSL2 is just 10% below Ubuntu