GeForce GTX980 time out issue

Accelerated Computing CUDA CUDA Programming and Performance
1

Problem: I cannot run tasks on GeForce GTX980 longer than few seconds. Here is a more detailed explanation. On a MS Windows computer (Windows 8.1, 64-bit, 16GB of RAM), I have GeForce GTX980 and an integrated card (Intel HD Graphics P4600/P4700). I am presently testing GPU performance for our algorithms. I started with a matrix-matrix multiplication. Every time I run this task I get “Display driver stopped responding and has recovered” message after few seconds of run time.

I disabled a watchdog from BIOS and set WDDM TDR enabled to False under Microsoft Display Driver in the NVIDIA Nsight Options. I only use Intel HD Graphics P4600/P4700 for display. I also get time out error if I run program with cuda-memcheck: Program hit CUDA_ERROR_LAUNCH_TIMEOUT (error 702) due to “the launch timed out and was terminated”

I even tried a simpler job:

“
global void func()
{
while(true){}
}
“
Same error came up for this simple kernel again after a few seconds.

This is very disturbing. I need to run calculations for much longer time without interruption. Some of optimization problems (should be good under GPU) can run for hours.

Please help.

Below is the device query of my GeForce GTX980:
C:\ProgramData\NVIDIA Corporation\CUDA Samples\v6.5\bin\win32\Debug\deviceQuery.exe Starting…

CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: “GeForce GTX 980”
CUDA Driver Version / Runtime Version 6.5 / 6.5
CUDA Capability Major/Minor version number: 5.2
Total amount of global memory: 4096 MBytes (4294967295 bytes)
(16) Multiprocessors, (128) CUDA Cores/MP: 2048 CUDA Cores
GPU Clock rate: 1342 MHz (1.34 GHz)
Memory Clock rate: 3505 Mhz
Memory Bus Width: 256-bit
L2 Cache Size: 2097152 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096,
4096)
Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 2 copy engine(s)
Run time limit on kernels: No
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
CUDA Device Driver Mode (TCC or WDDM): WDDM (Windows Display Driver Model)
Device supports Unified Addressing (UVA): No
Device PCI Bus ID / PCI location ID: 1 / 0
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 6.5, CUDA Runtime Version = 6.5,
NumDevs = 1, Device0 = GeForce
GTX 980
Result = PASS

2

you need to completely restart the computer in order for the registry changes to be valid. If you are using a laptop (GTX 980m), do not just turn it off, restart from the power bar.

3

If you plan to run a particular kernel for hours on the GPU, I would seriously suggest looking into the use of a separate GPU dedicated to compute tasks, preferably a Tesla GPU with ECC.

4

Just out of curiosity, would there be a need for ECC if the kernel was operating solely on integers (32 or 64 bit)?

At my firm we use both the Teslas and the GTX line. The Teslas for applications which use RDMA or for consumer medical products, and the GTX line for development.

For a current research project we are using a combination of the two actually (Tesla and GTX Titan Black), which has worked out well so far.

5

ECC on Tesla GPUs (also on some Quadro GPUs, I think) ensures the integrity of global memory contents, independent of what kind of data is stored there.

Personally I would very much hope that CUDA users deploying GPUs in medical products use GPUs with ECC, to avoid an “oh, I guess that was not a cancerous lesion in the image, just a flipped bit in GPU memory” effect.

Using consumer cards during the development process seems perfectly fine to me, but for anything mission critical I would recommend using a professional GPU. I understand that opinions on this matter differ, and the above is just my personal take. Personally, I use Quadros in my own PCs at home.

6

Funny enough, the Tesla-Titan research system will be used on pig subjects, and we will make sure those piggies get all the benefits of ECC. ;)

7

That will create good karma :-)

8

I actually disable watch dog from BIOS and I did restart my desktop right after it since simply changed registry key didn’t work for me even if I restarted it right after it. What else could be the reason? What operating system you use if that works for you?

9

I have never disabled the watchdog from the BIOS, rather I followed the instructions in the above posted video, and that has worked for me every time with Windows 7 SP1 64 bit and Windows 8.1 64 bit.

Did you try the exact instructions given in that video?

10

Oh! It works now! It did not work before if only change registry key. I need a combination of setting WDDM TDR enabled to False under Microsoft Display Driver in the NVIDIA Nsight Options, connect only Intel graphics card to display, change registry keys and disable watchdog from BIOS.