5-second limit in Windows XP Does it still apply with two cards?

Accelerated Computing CUDA CUDA Programming and Performance
1

Regaring the dreaded 5-second-kernel limit in Windows XP:

[url=“http://forums.nvidia.com/index.php?showtopic=29567&view=findpost&p=176998”]The Official NVIDIA Forums | NVIDIA

I am using a GeForce 5200 (PCI) as my display adapter, and the 8800 GTS as the GPU. The Windows desktop is NOT extended to the 8800.

While this prevents the maching from hanging on kernel runs longer than 5 seconds, such kernels return with an “unspecified launch failure.”

Questions for the panel:

Is it possible to run kernels longer than 5 seconds under Windows XP – say, with two PCIe cards and a different motherboard? Has anyone got it to work? (If so, what’s your configuration?)

If it’s not possible, is this limit permanent for Windows – one card, two or more cards, whatever?

2

This problem appears to have been fixed in v1.0 of CUDA and drivers.

3

If this problem is still relevant for anyone, here’s the ultimate solution:

http://msdn2.microsoft.com/en-us/library/ms797855.aspx

B)

4

Sorry for the confusion. The problem is NOT solved by the above. The machine doesn’t hang, but the kernel launch reports timeouts and doesn’t start. However now it fails on 7 seconds… :(

5

I wish there was more information on this problem. I’m new to CUDA programming but this problem could end my attempts at it. Does this mean that a CUDA program cannot last for more than 5 seconds or that an individual kernel launch cannot run on the machine for more than 5 seconds.

For example, if my program is executing a fluid solver and my kernel updates the flow field through one time step, does this mean that as long as the kernel can update the flow field in under 5 seconds I can simply re-run the kernel through as many time steps as I want. OR does it mean that my whole simulation has to finish in under 5 seconds!

Thanks for any help.

6

Each individual kernel launch must be under 5 seconds. I wouldn’t worry about it for you application. Each timestep is likely to take only a few tens of microseconds. If you make your simulation big enough to need 5 seconds of processing time, you will run out of memory on the device.

7

I was pretty sure that was the case (anything else wouldn’t have made much sense) but also worried enough to clarify. Thanks for the help.