Mon Nov 4 10:28:00 2013 | | OpenCL: NVIDIA GPU 0: GeForce GT 750M (driver version 8.20.15 310.40.15f03, device version OpenCL 1.2, 2048MB, 2048MB available, 30 GFLOPS peak)
Mon Nov 4 10:28:00 2013 | | OpenCL: Intel GPU 0: Iris Pro (driver version 1.2(Sep 29 2013 19:45:54), device version OpenCL 1.2, 1024MB, 1024MB available, 2688 GFLOPS peak)
OpenCL support but no CUDA support.
As you noted in the other thread, that's indeed a strange driver version. Our 5.5.28 driver shows the driver as being "8.18.27 310.40.05f01". I haven't seen any reference to what CUDA version he's got installed. But yes, technically the GT 750M (Kepler) should support up to sm_35. According to NVIDIA that should work...
Quote:
Apropos, seeing how that user has a correctly detected Intel GPU, are there plans to release an Intel_OpenCL_GPU application for the Mac?
Not yet, but that's a good point, in particular given the new MBPs (seems to be what the other thread's OP uses). We'll look into it, shouldn't be too hard to do. I think we're going to deploy a binary on albert soonish...
As you noted in the other thread, that's indeed a strange driver version. Our 5.5.28 driver shows the driver as being "8.18.27 310.40.05f01". I haven't seen any reference to what CUDA version he's got installed.
Well, that's the problem with the present available CUDA drivers, they don't work on Mavericks. And that's why his BOINC shows no CUDA capable GPU detected.
As for the weird driver numbering, I asked Charlie about that. He told me: They have two drivers on Macs which are installed together. One is the CUDA driver, which is version 5.5.28 in this case. The other is the general NVIDIA GPU driver, which has always had a long string of numbers. I have no idea what the numbers mean.
A correctly detected GPU on the Mac shows up like this:
Quote:
24-Apr-2013 08:40:37 [---] NVIDIA GPU 0: GeForce GT 650M (driver version 5.0.45, CUDA version 5.0, compute capability 3.0, 1024MB, 693MB available, 691 GFLOPS peak)
24-Apr-2013 08:40:37 [---] OpenCL: NVIDIA GPU 0: GeForce GT 650M (driver version CLH 1.0, device version OpenCL 1.1, 1024MB, 693MB available)
Although, checking the thread the above messages came from... the non-CUDA detection for the earlier Mac should yield the message "BOINC does not support CUDA on this computer." and it doesn't. Hmmm.
Well, that's the problem with the present available CUDA drivers, they don't work on Mavericks. And that's why his BOINC shows no CUDA capable GPU detected.
That shouldn't be the case as I described earlier. CUDA 5.5.28 officially supports 10.9, just not the old sm_1x GPUs. If it doesn't work it's a bug. I'll update my bug report at NVIDIA and let them know...
Quote:
As for the weird driver numbering, I asked Charlie about that. He told me: They have two drivers on Macs which are installed together. One is the CUDA driver, which is version 5.5.28 in this case. The other is the general NVIDIA GPU driver, which has always had a long string of numbers. I have no idea what the numbers mean.
This isn't really correct. 5.5.28 is the CUDA release which includes a driver which has the build number I quoted. The driver's build number can be seen in the CUDA system preference pane on OS X as well as in BOINC's OpenCL output.
Quote:
A correctly detected GPU on the Mac shows up like this:
Quote:
24-Apr-2013 08:40:37 [---] NVIDIA GPU 0: GeForce GT 650M (driver version 5.0.45, CUDA version 5.0, compute capability 3.0, 1024MB, 693MB available, 691 GFLOPS peak)
24-Apr-2013 08:40:37 [---] OpenCL: NVIDIA GPU 0: GeForce GT 650M (driver version CLH 1.0, device version OpenCL 1.1, 1024MB, 693MB available)
You seem to be using an older BOINC client. 7.2.23 does list the driver version as described above:
Thu Nov 7 10:44:05 2013 | | OpenCL: NVIDIA GPU 0: GeForce GT 120 (driver version 8.18.27 310.40.05f01, device version OpenCL 1.0, 512MB, 512MB available, 90 GFLOPS peak)
Thu Nov 7 10:44:05 2013 | | OpenCL: NVIDIA GPU 1: GeForce GTX 285 (driver version 8.18.27 310.40.05f01, device version OpenCL 1.0, 1024MB, 1024MB available, 708 GFLOPS peak)
Note: it's expected that these GPUs aren't detected as CUDA devices since they're sm_1x GPUs.
The question here is the following: why has Richard's machine a higher/younger driver build number than the latest official 5.5.28 CUDA release? That's why I was talking about the missing CUDA version reference in his post - I have no idea where his driver came from.
I was using all examples lifted from the BOINC forums. I don't have a computer with an Nvidia card, nor a Mac. :-)
Quote:
The question here is the following: why has Richard's machine a higher/younger driver build number than the latest official 5.5.28 CUDA release? That's why I was talking about the missing CUDA version reference in his post - I have no idea where his driver came from.
Might also be a difference in how Apple allows vendors to show the information in their drivers, from how it's done under other platforms. How many Linux distros aren't there that show (unknown version), when the Windows drivers can show that correct bit of info?
driver version required max: -29053, supplied: 33140
Boinc logs say,
07/11/2013 11:21:35 | | OpenCL: Intel GPU 0: Intel(R) HD Graphics 4000 (driver version 9.18.10.3071, device version OpenCL 1.2, 1496MB, 1496MB available, 45 GFLOPS peak)
Is the maximum driver version supported Intel 29053? If so, do you have a link; I can only find newer drivers?
I did roll back the driver from 33140 to 3071. There seems to be some confusion within Boinc about that. Note that I was using Beta versions of Boinc, so maybe a bug, but now on 7.0.64 (x64).
2013-11-07 11:23:37.6953 [PID=13081] [version] Checking plan class 'BRP4G-cuda32'
2013-11-07 11:23:37.6953 [PID=13081] [version] parsed project prefs setting 'gpu_util_brp': 1.000000
2013-11-07 11:23:37.6953 [PID=13081] [version] driver version required max: -29053, supplied: 33140
- so that's your CUDA driver version.
Check the preference screenshots I posted in message 127634 - those are the minimum required, though you can add others for other devices if you want. At the moment, the server isn't even checking the plan_class for 'opencl-intel_gpu', so a preference must be set wrong - compare my log at http://einstein.phys.uwm.edu/host_sched_logs/8864/8864187
Thanks Richard, that's it working again on the i7 system. Was thinking it might have been something to do with using Bam! and profiles, or just 7.2.x. Just had to select,
Run only the selected applications
Binary Radio Pulsar Search (Arecibo): yes
Binary Radio Pulsar Search (Arecibo, GPU): yes
Binary Radio Pulsar Search (Perseus Arm Survey): no
Gravitational Wave S6 Directed Search (CasA) : no
Gamma-ray pulsar search #2: no
Run CPU versions of applications for which GPU versions are available no
Also got it going on a laptop with the same configuration, but first had to update the driver,
Automatic driver update didn't shift it from 9.17.10.2867. So manually updated to 3071. Odd that W7 updates automatically to 3071 but W8 says the driver in use is up to date, when it's too old to work here.
The question here is the following: why has Richard's machine a higher/younger driver build number than the latest official 5.5.28 CUDA release?
For the sake of completeness: the slightly different driver versions represent Apple's own versioning for different hardware models. New models like the MBP with the GT 750M ship with "310.40.15f03" while older ones shipped with "310.40.05f01". That seemingly older version doesn't have to change, even when the older hardware is upgraded to the same OS X 10.9 build - it's effectively tied to the hardware.
Bottom line: if you want to run CUDA on OS X, you need to manually install the CUDA driver. In case of BOINC and E@H the driver alone is sufficient, no CUDA Toolkit or samples required.
RE: FYI: CUDA 5.5.28 is
)
Which then gives things like this:
Mon Nov 4 10:28:00 2013 | | OpenCL: NVIDIA GPU 0: GeForce GT 750M (driver version 8.20.15 310.40.15f03, device version OpenCL 1.2, 2048MB, 2048MB available, 30 GFLOPS peak)
Mon Nov 4 10:28:00 2013 | | OpenCL: Intel GPU 0: Iris Pro (driver version 1.2(Sep 29 2013 19:45:54), device version OpenCL 1.2, 1024MB, 1024MB available, 2688 GFLOPS peak)
OpenCL support but no CUDA support.
Apropos, seeing how that user has a correctly detected Intel GPU, are there plans to release an Intel_OpenCL_GPU application for the Mac?
Edit... totally off topic: I see that my RAC at the time of writing is in accordance with the driver version: RAC 5,528 ;-)
Hi
)
Hi Jord,
As you noted in the other thread, that's indeed a strange driver version. Our 5.5.28 driver shows the driver as being "8.18.27 310.40.05f01". I haven't seen any reference to what CUDA version he's got installed. But yes, technically the GT 750M (Kepler) should support up to sm_35. According to NVIDIA that should work...
Not yet, but that's a good point, in particular given the new MBPs (seems to be what the other thread's OP uses). We'll look into it, shouldn't be too hard to do. I think we're going to deploy a binary on albert soonish...
Best,
Oliver
Einstein@Home Project
RE: As you noted in the
)
Well, that's the problem with the present available CUDA drivers, they don't work on Mavericks. And that's why his BOINC shows no CUDA capable GPU detected.
As for the weird driver numbering, I asked Charlie about that. He told me:
They have two drivers on Macs which are installed together. One is the CUDA driver, which is version 5.5.28 in this case. The other is the general NVIDIA GPU driver, which has always had a long string of numbers. I have no idea what the numbers mean.
A correctly detected GPU on the Mac shows up like this:
Although, checking the thread the above messages came from... the non-CUDA detection for the earlier Mac should yield the message "BOINC does not support CUDA on this computer." and it doesn't. Hmmm.
RE: Well, that's the
)
That shouldn't be the case as I described earlier. CUDA 5.5.28 officially supports 10.9, just not the old sm_1x GPUs. If it doesn't work it's a bug. I'll update my bug report at NVIDIA and let them know...
This isn't really correct. 5.5.28 is the CUDA release which includes a driver which has the build number I quoted. The driver's build number can be seen in the CUDA system preference pane on OS X as well as in BOINC's OpenCL output.
You seem to be using an older BOINC client. 7.2.23 does list the driver version as described above:
Note: it's expected that these GPUs aren't detected as CUDA devices since they're sm_1x GPUs.
The question here is the following: why has Richard's machine a higher/younger driver build number than the latest official 5.5.28 CUDA release? That's why I was talking about the missing CUDA version reference in his post - I have no idea where his driver came from.
Oliver
Einstein@Home Project
RE: You seem to be using an
)
I was using all examples lifted from the BOINC forums. I don't have a computer with an Nvidia card, nor a Mac. :-)
Might also be a difference in how Apple allows vendors to show the information in their drivers, from how it's done under other platforms. How many Linux distros aren't there that show (unknown version), when the Windows drivers can show that correct bit of info?
Can't get Einstein GPU work
)
Can't get Einstein GPU work for the Intel iGPU (HD4000) on my i7-3770K.
http://einstein.phys.uwm.edu//host_sched_logs/6333/6333837
Note this line,
Boinc logs say,
07/11/2013 11:21:35 | | OpenCL: Intel GPU 0: Intel(R) HD Graphics 4000 (driver version 9.18.10.3071, device version OpenCL 1.2, 1496MB, 1496MB available, 45 GFLOPS peak)
Is the maximum driver version supported Intel 29053? If so, do you have a link; I can only find newer drivers?
I did roll back the driver from 33140 to 3071. There seems to be some confusion within Boinc about that. Note that I was using Beta versions of Boinc, so maybe a bug, but now on 7.0.64 (x64).
Thanks,
The line you are quoting is
)
The line you are quoting is in the context:
- so that's your CUDA driver version.
Check the preference screenshots I posted in message 127634 - those are the minimum required, though you can add others for other devices if you want. At the moment, the server isn't even checking the plan_class for 'opencl-intel_gpu', so a preference must be set wrong - compare my log at http://einstein.phys.uwm.edu/host_sched_logs/8864/8864187
Thanks Richard, that's it
)
Thanks Richard, that's it working again on the i7 system. Was thinking it might have been something to do with using Bam! and profiles, or just 7.2.x. Just had to select,
-
Binary Radio Pulsar Search (Arecibo): yes
Run CPU versions of applications for which GPU versions are available noBinary Radio Pulsar Search (Arecibo, GPU): yes
Binary Radio Pulsar Search (Perseus Arm Survey): no
Gravitational Wave S6 Directed Search (CasA) : no
Gamma-ray pulsar search #2: no
Also got it going on a laptop with the same configuration, but first had to update the driver,
http://einstein.phys.uwm.edu/host_sched_logs/9178/9178197
Automatic driver update didn't shift it from 9.17.10.2867. So manually updated to 3071. Odd that W7 updates automatically to 3071 but W8 says the driver in use is up to date, when it's too old to work here.
Glad you got it working :-)
)
Glad you got it working :-)
Einstein@Home Project
RE: The question here is
)
For the sake of completeness: the slightly different driver versions represent Apple's own versioning for different hardware models. New models like the MBP with the GT 750M ship with "310.40.15f03" while older ones shipped with "310.40.05f01". That seemingly older version doesn't have to change, even when the older hardware is upgraded to the same OS X 10.9 build - it's effectively tied to the hardware.
Also, Richard's 10.9 CUDA issue can now be explained and should be easily resolvable...
Bottom line: if you want to run CUDA on OS X, you need to manually install the CUDA driver. In case of BOINC and E@H the driver alone is sufficient, no CUDA Toolkit or samples required.
Best,
Oliver
Einstein@Home Project