Simultaneous GPU CPU Use

taylorg

After installing an GT 730 x1 video card with the latest NVidia 466.11 and 466.27 "game ready" drivers, my Rapid Miner processes with standard modeling operators (not from extensions) appear to run on both the CPU and GPU simultaneously while the RM window is open.  A screenshot of the javaw.exe process is attached that shows it using about 50% of both the CPU and the GPU.  This occurs for virtually every RM process, even the tutorial processes.

When the RM window is minimized, the GPU is no longer utilized but when the window is reopened, the GPU is again utilized at the same rate as before the window was minimized.  The GPU utilization varies between 15% and 80% and depends on the modeling operator and its parameter settings.  The output is the same whether the GPU is running or not.  The modeling intensive processes may run slightly faster with both CPU and GPU but it's not clearly speeding things up (the GPU has 1GB DDR3 at 900MHz so it's very slow). 

However, some scoring intensive processes (Naive Bayes with 100 kernels, 5000 rows, 400 regular numeric attributes, 1 polynomial label) seem to run nearly twice as fast when both CPU and GPU are utilized (200 cpu seconds  vs 100 cpu seconds.)  At least in these cases, I think I'm seeing useful behavior and not a rogue issue caused by a new graphics driver.

The setup includes RM Studio Educational 9.9.000 on a Dell Inspiron 580 i7-870 with the latest Windows 10 Home Edition.

I also have a Lenovo P50 with Quadro M2000M graphics and Win 10 but cannot replicate this behavior using the same RM processes, even after updating the drivers to 466.11.  It's probably not the exact same driver as used on the GeForce 730 above since the M2000M wasn't designed for gaming.  RM processes never use either the integrated graphics or the M2000M GPU on the P50.

Question 1:  When I see RM's javaw.exe using both CPU and GPU heavily, are useful RM tasks being offloaded from the CPU to the GPU?

Question 2:  If this is useful behavior, what could I do to replicate it on the P50 with a better GPU?

Question 3: If this is useful, why doesn't it work when the RM window is minimized?  I suspect this is caused by the NVidia driver seeing the RM process as a video game which wouldn't need GPU help when minimized.

jwpfau

Hi Taylor,

There are specific operators which can use the GPU, but in general it's only used for rendering the UI.

Having a dedicated graphics card installed also frees up the shared memory, maybe that's what's making the difference?

Greetings,
Jonas

taylorg

Hi Jonas

My comparisons involve running the same scoring routine (Apply Model) with the RM UI window minimized vs. open.  When the window is open, the GPU  used and the scoring log from the Apply Model operator shows half the CPU usage.  When the RM UI window is miniimized, there is no GPU usage related to javaw.exe (RM) and the scoring log shows twice the CPU execution times.  My other machine never shows any GPU usage from javaw.exe.

I think the issue relates to way the new NVidia drivers assign certain types of mathematical computations to the GPU that it thinks are game related.  The RM user interface takes very little GPU on other machines but I've seen javaw.exe taking 15-80% of GPU since I installed the new drivers on the x1 video card. 

If you run RM on Windows 10, open "Task Manager," click "details", open columns that display both CPU and GPU usage, do you see javaw.exe taking any GPU?  If so, is it ever more than 15%?  My two machines are vastly different:  One shows zero GPU usage and the other varies between 15-80%.