Attaching the debugger with Nsight and a Tesla K20 - not working at all for me.

Hello, all.

I had a server with Windows Server 2008 R2, 4 Tesla C2070s, and a GeForce 8800 to satisfy the requirement of a card to utilize for the debugger under Windows.

I received several K20s, and we replaced one of the C2070s with a K20, uninstalled the drivers, Nsight, CUDA Toolkit, and reinstalled with the Nsight 3.0 preview and the 306.94 drivers. To our surprise, despite explicitly advertising “This new release officially supports the new Kepler GK110 architecture found in Tesla® K20 and CUDA 5.0.”, the 306.94 drivers do not, in fact, support any variant of the Tesla K20. I went to look at explicitly attaching the driver to the unknown device in Device Manager, but found there were no entries for the K20 - the K10 was the only remotely relevant entry.

Okay, so I tried grabbing the suggested release for the K20 from the drivers page on the main site - 310.90 as of this writing. Installed that, it happily recognized the K20 and the C2070s, but not the 8800 - after all, it was the Quadro/NVS/Tesla driver, not the GeForce driver.

So I swapped the 8800 for an NVS 295.

All 4 Teslas, and the NVS 295, were seen by the OS.

However, none of the cards can be used for the Nsight debugger - attempting to doso causes the debugged application to crash under Visual Studio, and that’s the end of that.

I find it extraordinarily hard to believe that it’s not presently possible to use the Nsight debugger with the Tesla K20, NVIDIA’s pride and joy - what am I overlooking?

Did you happen to re-install the driver after you put in the NVS 295 (is it still 310.90)?

Are you able to run the application OK outside of Nsight? Any extra information on where it crashes? Could you try to debug any of the samples that come with Nsight, or the SDK?

Any information on this, and the application you are using would be helpful.

We have been testing the 310.90 driver.


I can give a bit of followup to the OP’s problem description:

We have a CUDA application that runs without error on the K20 – without the debugger – when compiled either for compute capability 2.0 or compute capability 3.0. But when debugging using NSight for Visual Studio, it bombs with a generic Windows application-failure dialog shortly after Visual Studio attaches to the NSight monitor. The app isn’t debuggable inside NSight at its failure point, so we can’t tell you where it fails.

Anyway, to stay on topic: Despite NVidia’s advertising to the contrary, we cannot make the developer driver v306.94 recognize the K20. Driver version 310.90 does recognize the K20 but appears to be incompatible with NSight for Visual Studio (v3.0.0.12350). So if somebody out there is actually using NSight Visual Studio to debug CUDA code on a K20, could you please let us know what driver version and NSight version you are using?



What about debugging the CUDA SDK samples or the samples that are shipped with Nsight? Do those work?

I can confirm that we have a win2008R2 server running the 310.90-quadro-tesla-winserv2008-2008re driver which works with nsight(VS2010) both locally and remotely. We have a C2050 and K20 on board. HOpe that helps.

oops should also have said we are using the same version of nsight

Would you mind trying the new Nsight Visual Studio Edition 3.0 RC1 release from the Early Access site?