I tried the new nvidia-drivers 325.15 with the same results.
My estimations for the performance of 325.15 have shown that it is ~6 times slower than 313.30 in EaH.
I also tried the new nvidia-drivers (319.49 and 325.15) with the new linux kernel (3.10.7) with the same results.
The 313.30 is the newest version of nvidia-drivers with stable performance on 3.8 and 3.10 kernels.
For the new linux kernels (> 3.8) nvidia-drivers need a patch. Gentoo only apply this patch to specific versions of nvidia-drivers.
If you would like to run 3.10 linux kernel with 313.30 nvidia-drivers you must apply manually the patch (link):
I'm seeing some performance increases under Linux with the latest 331 beta driver, however I had to make other changes to that particular PC at the same time (memory problems), so I would be interested in hearing from others confirming/rejecting this observation.
I'm seeing some performance increases under Linux with the latest 331 beta driver, however I had to make other changes to that particular PC at the same time (memory problems), so I would be interested in hearing from others confirming/rejecting this observation.
Cheers
HBE
Finally, a nvidia driver which has no performance issues with EaH !!!
It seems to be more efficient than 313.30 but I will wait for full wu processing from 331.13 driver. I will post the results...
This could have some important implications for E@H: If this new driver (when finally released) should really work well with our code, we would finally be in a position to provide BRP apps that are compiled against CUDA 5.5 runtine and linked with the CUDA 5.5 FFT lib, something that is high on our wish list. In theory this should also help performance.
Assuming that the 331 drivers work out is there any "special" consideration for which distro of Linux to use? Or should they drop in for any distro? I don't seem to recall anyone who has experienced improvement with these beta drivers mentioning the Linux distro on which they installed them.
Looking at the recent phoronix benchmark displaying compute performance of AMD versus NVIDIA, I would highly question anyone who states they are getting good performance with NVIDIA on Linux. Seems only AMD GPUs benefit from Linux right now.
From my first results I noticed an average decrease of 15% and 11,5% on the processing time for the Perseus Arm Survey and Arecibo WUs respectively, comparing the new nvidia-driver-331.13 and the nvidia-driver-313.30.
My system has an NVIDIA GTX-650 and an AMD PhenomII X6 1090T with 8GB DDR2-1066.
I am running Gentoo Linux (AMD64) with kernel 3.10.7 (custom kernel configuration).
The benchmark you quoted compares OpenCL on NVIDIA to OpenCL on AMD, which is another story. For E@H the relevant comparison would be a CUDA (version 3.2 at the moment) implementation on NVIDIA compared to an OpenCL implementation of the same problem on AMD. But this thread is not about AMD vs NVIDIA but about performance under different NVIDIA drivers, so let's continue this discussion here
@BackGroundMAN
Yup, that's about what I see on my test system (w/ a GTX 650 Ti) . I think this is quite remarkable, I hope it stays like this until the release version!
Thanks again for testing, results with other hardware are welcome, or course.
RE: Magic, this problem
)
Yes nvidia magic !!!!
I tried the new
)
I tried the new nvidia-drivers 325.15 with the same results.
My estimations for the performance of 325.15 have shown that it is ~6 times slower than 313.30 in EaH.
I also tried the new
)
I also tried the new nvidia-drivers (319.49 and 325.15) with the new linux kernel (3.10.7) with the same results.
The 313.30 is the newest version of nvidia-drivers with stable performance on 3.8 and 3.10 kernels.
For the new linux kernels (> 3.8) nvidia-drivers need a patch. Gentoo only apply this patch to specific versions of nvidia-drivers.
If you would like to run 3.10 linux kernel with 313.30 nvidia-drivers you must apply manually the patch (link):
wget https://bugzilla.kernel.org/attachment.cgi?id=102881 -O /etc/portage/patches/x11-drivers/nvidia-drivers-313.30/linux_3.10.patch
emerge -qv nvidia-drivers
Hi! I'm seeing some
)
Hi!
I'm seeing some performance increases under Linux with the latest 331 beta driver, however I had to make other changes to that particular PC at the same time (memory problems), so I would be interested in hearing from others confirming/rejecting this observation.
Cheers
HBE
RE: Hi! I'm seeing some
)
Finally, a nvidia driver which has no performance issues with EaH !!!
It seems to be more efficient than 313.30 but I will wait for full wu processing from 331.13 driver. I will post the results...
Hi! Thanks for testing,
)
Hi!
Thanks for testing, standing by for results :-).
This could have some important implications for E@H: If this new driver (when finally released) should really work well with our code, we would finally be in a position to provide BRP apps that are compiled against CUDA 5.5 runtine and linked with the CUDA 5.5 FFT lib, something that is high on our wish list. In theory this should also help performance.
Cheers
HB
Assuming that the 331 drivers
)
Assuming that the 331 drivers work out is there any "special" consideration for which distro of Linux to use? Or should they drop in for any distro? I don't seem to recall anyone who has experienced improvement with these beta drivers mentioning the Linux distro on which they installed them.
Looking at the recent
)
Looking at the recent phoronix benchmark displaying compute performance of AMD versus NVIDIA, I would highly question anyone who states they are getting good performance with NVIDIA on Linux. Seems only AMD GPUs benefit from Linux right now.
http://www.phoronix.com/scan.php?page=article&item=nvidia_geforce_gtx650&num=8
From my first results I
)
From my first results I noticed an average decrease of 15% and 11,5% on the processing time for the Perseus Arm Survey and Arecibo WUs respectively, comparing the new nvidia-driver-331.13 and the nvidia-driver-313.30.
My system has an NVIDIA GTX-650 and an AMD PhenomII X6 1090T with 8GB DDR2-1066.
I am running Gentoo Linux (AMD64) with kernel 3.10.7 (custom kernel configuration).
@mmstick : The benchmark
)
@mmstick :
The benchmark you quoted compares OpenCL on NVIDIA to OpenCL on AMD, which is another story. For E@H the relevant comparison would be a CUDA (version 3.2 at the moment) implementation on NVIDIA compared to an OpenCL implementation of the same problem on AMD. But this thread is not about AMD vs NVIDIA but about performance under different NVIDIA drivers, so let's continue this discussion here
@BackGroundMAN
Yup, that's about what I see on my test system (w/ a GTX 650 Ti) . I think this is quite remarkable, I hope it stays like this until the release version!
Thanks again for testing, results with other hardware are welcome, or course.
Cheers
HBE