I just recently built a new workstation, installed Linux Mint on it and just completed installation of Houdini. My setup is as follows:

2x Intel E5 2660
48gb Ram
Quadro 4000 & Tesla c2075

I'm having difficulties seeing if my tesla card is actually working the way it's supposed to in Houdini. This is what I get in the About tab:

OpenGL Vendor: NVIDIA Corporation
OpenGL Renderer: Quadro 4000/PCIe/SSE2
OpenGL Version: 4.2.0 NVIDIA 295.40
OpenGL Shading Language: 4.20 NVIDIA via Cg compiler
Detected: NVidia Professional
2048 MB
295.40.0.0

It does not show the Tesla card at all, although the Nvidia X Server settings shows it as the second gpu. Is there a terminal command that must be ran to enable the functionality between the quadro and tesla?

Houdini doesn't display OpenCL information in the About Houdini dialog, only OpenGL. Since you have Nvidia drivers > 275, Nvidia's Maximus tech should be automatically routing your sim to the compute card.

There are a couple of ways to ensure it's using the Telsa for compute:

1) try a really large volume sim, such as 400x400. The Quadro 4000 does not have enough memory to run this sim, while the Tesla does. If it's using the Telsa, you'd be able to simulate the volume over many frames.

2) (non-technical method) Put your hand by the exhaust vent of the Telsa while running a heavy sim. If it's computing, you should feel some pretty warm air flowing out.

Alternatively, there is the environment variable HOUDINI_OCL_DEVICENUMBER which you can set to ‘1’ to select the Tesla (zero is your primary display card).

Thanks for the response.

I went ahead and set the environment variable and confirmed through hconfig -a that it was set to the appropriate value of 1. I just started running a basic shelf tool flames sim and set the uniform div to 400 and made sure OpenCL was checked. Frame 38 took about 12 seconds to cook.

I am running the nvidia-smi terminal command and it shows the Quadro using 28% of its memory (566MB) and the tesla at 16% (878MB). This indicates that they're working (I think :? ) . I guess I was expecting a little more performance wise. Is there anything else I can check just to be sure?

Thanks for the help!

I just started running a basic shelf tool flames sim and set the uniform div to 400 and made sure OpenCL was checked. Frame 38 took about 12 seconds to cook. … I guess I was expecting a little more performance wise.

Well, to be fair, you are computing 64 million voxels in that case, with some subframe steps. It sounds like it's working based on the performance and memory stats; my Quadro 4000 tops out at around 320^3 before it runs out of VRAM.

You can try turning off the OpenCL toggle and see if it's slower – if the times are the same, chances are the CPU is crunching the numbers in both cases. There are some non-CL operations in DOPs as well, which can slow the sim down if you use them (not exactly sure which ones, offhand).

It's working. Running a smoke sim @ 400^3 and I've used about 62% of the memory. I'll try to optimize the scene more and push the GPU to the max.

Something very strange just started happening. I was running a simulation (PyroFX Volcano tutorial) and the resolution would increase/decrease every frame in the viewport, almost like a blurring effect on one frame and then back to the original resolution on the next frame. Driver issue? I tried switching the scene renderer, but that just made it worse….

-update-

My card is running a little hot (low 80s). That may be an issue, not sure.