Windows 10 simply isn't able to read GPU utilization while running these GPU apps (at least on Einstein or Albert). That is the situation also on desktop hosts and same for Nvidia and AMD cards. That visualizer might display other type of activity, I don't know what kind of. Probably it would show GPU activity when playing videos on social media websites, or connecting a mobile phone and uploading a cat video from it. Useless feature.
Standard 10 version keeps showing pretty much 0% utilization. Insider fast ring build says utilization is about 20%, while GPU-Z says utilization is 70% at the same time.
edit: yet more cleanup
* on build 17763 and Insider build's Task Manager there's also two new columns on the page where loaded apps, processes and columns for cpu, mem, disk, network and gpu are listed:
'Power usage' and 'Power usage trend'
A host with build 17763 and Nvidia GTX 960 running 1x Einstein opencl : 'Power usage' shows read color and says "Very high" for a running Einstein GPU process. At the same time the GPU Utilization of that process is about 0.5% (GPU-Z says it's more than 90%). Power usage trend shows nothing.
A host with insider build 18272 and AMD R9 270X running 1x Albert opencl: Power usage is "Moderate" while GPU usage is about 16% (GPU-Z says it's more than 70%). Power usage trend shows nothing.
Based on this H2G article, it looks like the numbers from task manager come from monitoring the OSes graphics layer and only attempt to measure actual graphics work. OTOH 65-85% task manager loads on my main personal PC don't appear to be impacting boinc's ability to run GPU work.
On the contrary, Windows is able to view the individual gpu utilization easily if you just use the provided Nvidia tool.
If you navigate to the C:\Program Files\NVIDIA Corporation\NVSMI directory and open a Terminal there, you can run
nvidia-smi and it will print out the Performance State, memory, temperature, power usage in watts and utilization on each installed nvidia card.
I believe MSI Afterburner, which works fo both AMD and Nvidia cards, shows the same thing and will even let you adjust things if you wish. I use it to adjust the fan speed to keep mine cool.
Task Manager's Performance tab does display computing as well but you need to look at the right engine. On my system that is Compute_0, other systems may have different engine that is of interest.
Task Manager's Performance tab does display computing as well but you need to look at the right engine. On my system that is Compute_0, other systems may have different engine that is of interest.
You are right. Thanks. I had never noticed there were pull-down "menus" to choose what those graphs should display.
It looks like for Nvidia the utilization goes well hand-in-hand with GPU-Z graph. For couple of AMD cards it seems to constantly show a little bit lower percentage than GPU-Z (numerical difference about 10).
Hi Szydas,
Welcome to the Einstein project. Thank you for joining!
szydas wrote:
25.05.2019 09:32:56 | | CUDA: NVIDIA GPU 0: GeForce RTX 2080 (driver version 430.39, CUDA version 10.1, compute capability 7.5, 4096MB, 3549MB available, 10068 GFLOPS peak)
This line from your event log shows that the CUDA capability of your GPU is recognised. The Einstein app uses OpenCL rather than CUDA and there is no indication that you have OpenCL libraries installed. I don't use Windows so have no direct experience but from what others have said, drivers supplied by MS don't include OpenCL libs. You may need to get the full drivers from nVidia in order to get the libs that are needed.
The tasks list for your computer shows lots of CPU tasks only - 209 of them - with no indication of any GPU tasks. Whilst your BOINC client requested tasks for NVIDIA GPU, none were sent, suggesting that the server knew your GPU didn't have the required OpenCL compute libs.
Apart from missing libs, another problem you may encounter is to do with the size of your cache of downloaded tasks. Whilst you do have some nice fast hardware, 209 CPU tasks are going to take quite a while to crunch. Sometimes when starting, estimates for crunch time are not very accurate so it's wise to deliberately reduce your work cache settings to something low (eg. 0.5 days worth) until you see how long tasks really do take. Once things settle, you can then increase to a value that suits your circumstances. A low cache setting is particularly important for when you first start crunching GPU tasks. The results of one type of task will affect the estimates for other types. Fast finishing GPU tasks will reduce the estimates for CPU tasks and this can lead to significant over-fetching of tasks for CPUs. This problem is magnified if you also have a multi-day work cache setting.
If you need any assistance with changing settings, please ask.
Thanks for your support. I've checked it, and I did not find OpenCl support for my GPU. That's strange. NVIDIA claims OpenCL support is included in CUDA libraries in every driver since years. I have up-to-date drivers from NVIDIA installed, but both GPU-Z and GPU Caps Viewer did not find any OpenCL drivers or GPU supporting it... Additionaly on internet there's no single tutorial how to enable it; everyone states it SHOULD be enabled by default in NVIDIA drivers. Kinda frustrating...
Another thought is maybe task
)
Another thought is maybe task manager is showing your IGPU usage is which should be zero since you have a real GPU, your 650m
Windows 10 simply isn't able
)
Windows 10 simply isn't able to read GPU utilization while running these GPU apps (at least on Einstein or Albert). That is the situation also on desktop hosts and same for Nvidia and AMD cards. That visualizer might display other type of activity, I don't know what kind of. Probably it would show GPU activity when playing videos on social media websites, or connecting a mobile phone and uploading a cat video from it. Useless feature.
Standard 10 version keeps showing pretty much 0% utilization. Insider fast ring build says utilization is about 20%, while GPU-Z says utilization is 70% at the same time.
edit: yet more cleanup
* on build 17763 and Insider build's Task Manager there's also two new columns on the page where loaded apps, processes and columns for cpu, mem, disk, network and gpu are listed:
'Power usage' and 'Power usage trend'
A host with build 17763 and Nvidia GTX 960 running 1x Einstein opencl : 'Power usage' shows read color and says "Very high" for a running Einstein GPU process. At the same time the GPU Utilization of that process is about 0.5% (GPU-Z says it's more than 90%). Power usage trend shows nothing.
A host with insider build 18272 and AMD R9 270X running 1x Albert opencl: Power usage is "Moderate" while GPU usage is about 16% (GPU-Z says it's more than 70%). Power usage trend shows nothing.
Based on this H2G article, it
)
Based on this H2G article, it looks like the numbers from task manager come from monitoring the OSes graphics layer and only attempt to measure actual graphics work. OTOH 65-85% task manager loads on my main personal PC don't appear to be impacting boinc's ability to run GPU work.
https://www.howtogeek.com/351073/how-to-monitor-gpu-usage-in-the-windows-task-manager/
On the contrary, Windows is
)
On the contrary, Windows is able to view the individual gpu utilization easily if you just use the provided Nvidia tool.
If you navigate to the C:\Program Files\NVIDIA Corporation\NVSMI directory and open a Terminal there, you can run
nvidia-smi and it will print out the Performance State, memory, temperature, power usage in watts and utilization on each installed nvidia card.
Keith Myers wrote:On the
)
I believe MSI Afterburner, which works fo both AMD and Nvidia cards, shows the same thing and will even let you adjust things if you wish. I use it to adjust the fan speed to keep mine cool.
Task Manager's Performance
)
Task Manager's Performance tab does display computing as well but you need to look at the right engine. On my system that is Compute_0, other systems may have different engine that is of interest.
Juha wrote:Task Manager's
)
You are right. Thanks. I had never noticed there were pull-down "menus" to choose what those graphs should display.
It looks like for Nvidia the utilization goes well hand-in-hand with GPU-Z graph. For couple of AMD cards it seems to constantly show a little bit lower percentage than GPU-Z (numerical difference about 10).
Similiar problem here. GPU-Z
)
Similiar problem here. GPU-Z (latest version) shows 0% load, GPU temperatureis also low.
25.05.2019 09:32:55 | | cc_config.xml not found - using defaults
25.05.2019 09:32:55 | | Starting BOINC client version 7.14.2 for windows_x86_64
25.05.2019 09:32:55 | | log flags: file_xfer, sched_ops, task
25.05.2019 09:32:55 | | Libraries: libcurl/7.47.1 OpenSSL/1.0.2g zlib/1.2.8
25.05.2019 09:32:55 | | Data directory: C:\ProgramData\BOINC
25.05.2019 09:32:55 | | Running under account szydas
25.05.2019 09:32:56 | | CUDA: NVIDIA GPU 0: GeForce RTX 2080 (driver version 430.39, CUDA version 10.1, compute capability 7.5, 4096MB, 3549MB available, 10068 GFLOPS peak)
25.05.2019 09:32:56 | | Host name: stach
25.05.2019 09:32:56 | | Processor: 16 GenuineIntel Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz [Family 6 Model 158 Stepping 12]
25.05.2019 09:32:56 | | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 fma cx16 sse4_1 sse4_2 movebe popcnt aes f16c rdrandsyscall nx lm avx avx2 vmx smx tm2 pbe fsgsbase bmi1 hle smep bmi2
25.05.2019 09:32:56 | | OS: Microsoft Windows 10: Professional x64 Edition, (10.00.17763.00)
25.05.2019 09:32:56 | | Memory: 31.92 GB physical, 36.67 GB virtual
25.05.2019 09:32:56 | | Disk: 465.16 GB total, 163.86 GB free
25.05.2019 09:32:56 | | Local time is UTC +2 hours
25.05.2019 09:32:56 | | No WSL found.
25.05.2019 09:32:56 | Einstein@Home | URL http://einstein.phys.uwm.edu/; Computer ID 12779640; resource share 100
25.05.2019 09:32:56 | Einstein@Home | General prefs: from Einstein@Home (last modified 24-May-2019 13:30:09)
25.05.2019 09:32:56 | Einstein@Home | Host location: none
25.05.2019 09:32:56 | Einstein@Home | General prefs: using your defaults
25.05.2019 09:32:56 | | Reading preferences override file
25.05.2019 09:32:56 | | Preferences:
25.05.2019 09:32:56 | | max memory usage when active: 16343.46 MB
25.05.2019 09:32:56 | | max memory usage when idle: 29418.23 MB
25.05.2019 09:32:56 | | max disk usage: 100.00 GB
25.05.2019 09:32:56 | | (to change preferences, visit a project web site or select Preferences in the Manager)
25.05.2019 09:32:56 | | Setting up project and slot directories
25.05.2019 09:32:56 | | Checking active tasks
25.05.2019 09:32:56 | | Setting up GUI RPC socket
25.05.2019 09:32:56 | | Checking presence of 326 project files
25.05.2019 09:33:02 | Einstein@Home | Sending scheduler request: To fetch work.
25.05.2019 09:33:02 | Einstein@Home | Requesting new tasks for CPU and NVIDIA GPU
25.05.2019 09:34:41 | Einstein@Home | Scheduler request completed: got 45 new tasks
Any hint how to force GPU processing?
Hi Szydas, Welcome to the
)
Hi Szydas,
Welcome to the Einstein project. Thank you for joining!
This line from your event log shows that the CUDA capability of your GPU is recognised. The Einstein app uses OpenCL rather than CUDA and there is no indication that you have OpenCL libraries installed. I don't use Windows so have no direct experience but from what others have said, drivers supplied by MS don't include OpenCL libs. You may need to get the full drivers from nVidia in order to get the libs that are needed.
The tasks list for your computer shows lots of CPU tasks only - 209 of them - with no indication of any GPU tasks. Whilst your BOINC client requested tasks for NVIDIA GPU, none were sent, suggesting that the server knew your GPU didn't have the required OpenCL compute libs.
Apart from missing libs, another problem you may encounter is to do with the size of your cache of downloaded tasks. Whilst you do have some nice fast hardware, 209 CPU tasks are going to take quite a while to crunch. Sometimes when starting, estimates for crunch time are not very accurate so it's wise to deliberately reduce your work cache settings to something low (eg. 0.5 days worth) until you see how long tasks really do take. Once things settle, you can then increase to a value that suits your circumstances. A low cache setting is particularly important for when you first start crunching GPU tasks. The results of one type of task will affect the estimates for other types. Fast finishing GPU tasks will reduce the estimates for CPU tasks and this can lead to significant over-fetching of tasks for CPUs. This problem is magnified if you also have a multi-day work cache setting.
If you need any assistance with changing settings, please ask.
Cheers,
Gary.
Thanks for your support. I've
)
Thanks for your support. I've checked it, and I did not find OpenCl support for my GPU. That's strange. NVIDIA claims OpenCL support is included in CUDA libraries in every driver since years. I have up-to-date drivers from NVIDIA installed, but both GPU-Z and GPU Caps Viewer did not find any OpenCL drivers or GPU supporting it... Additionaly on internet there's no single tutorial how to enable it; everyone states it SHOULD be enabled by default in NVIDIA drivers. Kinda frustrating...