Windows 10 + WDDM 2.0 impact on CUDA?

Is there anything new in Win10 and WDDM 2.0 that CUDA developers should be aware of?

For example, what’s the status of kernel launch overhead on WDDM 2.0 (vs. TCC and WDDM 1.x)?

A summary would be appreciated!

Has that particular issue improved with newer versions of CUDA (6.5 and up)? I have not seen much difference with kernel launch overhead when comparing the same application and hardware Windows 7 x64 vs Ubuntu.

Also interested in the answer to this question, and look forward to the TCC driver for the GTX Titan X.

Would be nice if someone here could give some info regarding the performance impact of wddm 2.0 (compared with wddm on windows 7 or 8), especially the kernel launch overhead.

Some subjective observations after a couple months of daily Win10+CUDA:

Win10, CUDA 7.5 and GL interop work well.

I haven’t noticed any performance issues.

Still no VS2015 support but I assume that’s coming in CUDA 8.

Also, I don’t see DX12 interop but am working with DX11 interop for now.

Unlike Win7, I haven’t yet managed to wedge the Windows UI with an errant CUDA app.

There is always tomorrow. :)

For your large allocation random memory read test is there any performance difference between Windows 10 WDDM 2.0 and Windows 7/8 WDDM 1.0 ?

This thread touched on the topic using your test code;

https://devtalk.nvidia.com/default/topic/878455/?comment=4683608

.

My results match @Genoil’s results:

Notice that the rolloff points are half that of Win7.

That is disappointing for Windows 10, guess I am stuck with 7/8.1 .

I mentioned this before, but Genoil said for Windows 8.1 the dropoff occurs at 1 GB for GK200 (which covers the GTX 980M I believe), but for my tests using Windows 8.1 on my laptop with a GTX 980M I see the same dropoff as Windows 7;

GeForce GTX 980M : 12 SM : 4096 MB
Probing from: 128 - 2560 MB ...
alloc MB, probe MB,    msecs,     GB/s
     128,    28672,   212.03,   132.06
     256,    28672,   216.34,   129.42
     384,    28672,   216.15,   129.54
     512,    28672,   222.36,   125.92
     640,    28672,   235.61,   118.84
     768,    28672,   244.99,   114.29
     896,    28672,   251.74,   111.23
    1024,    28672,   256.88,   109.00
    1152,    28672,   261.12,   107.23
    1280,    28672,   264.29,   105.94
    1408,    28672,   266.88,   104.91
    1536,    28672,   268.68,   104.21
    1664,    28672,   270.54,   103.50
    1792,    28672,   272.12,   102.89
    1920,    28672,   273.49,   102.38
    2048,    28672,   274.70,   101.93
    2176,    28672,   499.21,    56.09
    2304,    28672,   894.42,    31.31
    2432,    28672,  1271.57,    22.02
    2560,    28672,  1612.87,    17.36