Linux vs Windows GPU performance difference
13 Aug 2016 4:59:41 UTC
Topic 199953
(moderation:
I recently moved a Nvidia GTX 950 video card from a Linux computer to a Windows computer and the change in performance has surprised me. In the Linux computer, the median run time for Binary Radio Pulsar Search (Arecibo, GPU) v1.52 (BRP4G-Beta-cuda32-nv270) was 5,679.66 seconds. In the Windows computer, the median run time for Binary Radio Pulsar Search (Arecibo, GPU) v1.52 (BRP4G-Beta-cuda32-nv301) is 6,756.59 seconds. I am trying to understand why the run time has gone up by almost 19%.
The Linux computer has an Intel i7-4790k, uses Nvidia driver 352.79, and runs CPU tasks on 3 of the cores. The Windows computer has an Intel i5-4690, uses Nvidia driver 368.81, and is not running any CPU tasks. Both computers have 3 tasks running on the GPU. Since the Windows computer isn’t running any CPU tasks, I don’t think the small CPU speed difference is the reason for the change.
My best guess is either the driver version or the difference between the Linux and Windows application is the cause of the increased runtime. Or maybe there is something else I am not considering. Do any of you have any ideas?
Ryan
I guess you might need to
)
I guess you might need to change things around (swap processors) to get a better answer.
The two systems are different so it's not only Linux / Windows that is different.
http://ark.intel.com/compare/80807,80810
What motherboard and memory were you running in each?
n12365 wrote:The Windows
)
That driver implements CUDA 8. It has been shown that the earlier drivers that use CUDA 7.5 work better on Einstein (as well as GPUGrid). Try 359.06, which was the last of the CUDA 7.5 drivers. That should cut the difference with Linux to maybe 10% as a guess. But Linux will still be faster, because Windows (ever since Vista) uses the Windows Display Driver Model (WDDM), which includes some additional overhead for crash protection.
AgentB wrote:What motherboard
)
Both systems use a Gigabyte GA-B85M-DS3H-A motherboard (Intel B85 chipset). The Windows system has 16 gigabytes of memory and the Linux system has 8 gigabytes of memory.
Thanks Jim. I have a GTX
)
Thanks Jim. I have a GTX 1060 on order, so I will be forced to use one of the newer drivers with it. Looks like the 950 will be going back in the Linux machine. Maybe the project will release a CUDA 8 application to take advantage of the new Pascal GPUs.
n12365 wrote:I have a GTX
)
As your Windows machines are Windows 10, you may well find that running the GTX 1060 on one of them will give you a pretty substantial, steady memory leak. There is [url=https://einsteinathome.org/content/memory-depletion-graphics-driver-related]a thread [/url] on this topic in the Problems and Bug Reports forum.
You may find it useful to your own interests to check for the problem, and it would be helpful to us in understanding the scope of the problem to learn whether you suffer it.
archae86 wrote:As your
)
I was aware of the memory leak problem when I ordered the 1060. I will share my experience with the card, either positive or negative.
Your post in the Nvidia Pascal and AMD Polaris thread is the only reason I was aware of the problem. Thanks for cross posting the information.