What controls the amount of CPU time used for the GPU WUs?

I have a P100 in system running linux Mint 20.1 using the 470.86 driver and it uses practically a full CPU thread for each WU. Here's the host tasks...https://milkyway.cs.rpi.edu/milkyway/results.php?hostid=906244

I also have a Titan V running in a nearly identical system also linux Mint 20.1 using the 470.86 driver but only uses about 1/3 of a CPU thread. Why? I understand the differences in the 2 GPUs but what controls how much CPU is used?

I've also noticed the Titan V uses up to 1266MiB GPU memory per WU while the P100 only uses up to 1066MiB. Both GPUs have 12GB of memory.

Thanks,
scole of TSBT

What controls the amount of CPU time used for the GPU WUs?

I have a P100 in system running linux Mint 20.1 using the 470.86 driver and it uses practically a full CPU thread for each WU. Here's the host tasks...https://milkyway.cs.rpi.edu/milkyway/results.php?hostid=906244

I also have a Titan V running in a nearly identical system also linux Mint 20.1 using the 470.86 driver but only uses about 1/3 of a CPU thread. Why? I understand the differences in the 2 GPUs but what controls how much CPU is used?

I've also noticed the Titan V uses up to 1266MiB GPU memory per WU while the P100 only uses up to 1066MiB. Both GPUs have 12GB of memory.

Thanks,
scole of TSBT

I think different cards can do more of the work themselves, so the CPU has to do more or less work to help out. I also think the AMDs need less CPU assistance.

I have for example an AMD Radeon R9 280X on three different machines, all running Windows 10:
Ryzen R9 3900XT: 4 Milkyway tasks on the GPU, GPU fully utilised, 150 GPU MB per WU (yours is rather high!), 0.1% CPU usage = 2.4% of a core per WU.
Pentium N3700: 1 Milkyway task on the GPU (old GPU, crashes if I use too much RAM), GPU 70% utilised, 300 GPU MB per WU, 10% CPU usage = 40% of a core per WU.
Intel Core2 Q8400 CPU: 2 Milkyway tasks on the GPU (again, old GPU, crashes if given too much to do), GPU 90% utilized, 100 GPU MB per WU, 4% CPU usage = 16% of a core per WU.

Note the CPU usage is intermittent, peaking 4 times per WU, are you taking the average?

Your host looks very odd:
https://milkyway.cs.rpi.edu/milkyway/show_host_detail.php?hostid=906244
What's a 0000 CPU? And why does it have a weird number of cores (56)? "Genuine Intel(R) CPU 0000 @ 2.40GHz"
Why does your card report 12GB and 4GB on the same line? Yet when I look up a P100, it should have 16GB.

System has dual Xeon V4 engineering sample CPUs in it, 28 threads per CPU, 56 threads total for system.

Don't know why it reports 12GB and 4GB. As for 12GB instead of 16GB, there are both 12GB and 16GB model P100s. Something else I don't understand, if you look at the task info it says it found 2 CL devices. On WUs for all other devices I've only seen 1 CL device reported.

System has dual Xeon V4 engineering sample CPUs in it, 28 threads per CPU, 56 threads total for system.

Don't know why it reports 12GB and 4GB. As for 12GB instead of 16GB, there are both 12GB and 16GB model P100s. Something else I don't understand, if you look at the task info it says it found 2 CL devices. On WUs for all other devices I've only seen 1 CL device reported.

I didn't realise they made CPUs with 28 cores. But then I suppose I've seen 6 and 24, so not always powers of 2.

Is your card a dual processor GPU? Like the AMD 7990?

If it does have multiple devices, it's not like a 7990 which actually appeared as 2 GPUs to your computer and BOINC. I think the K80 tesla is like that. I think the P100 and later teslas can be partitioned in a way so they can be shared in a data center environment but I'm not sure how that's done.

System has dual Xeon V4 engineering sample CPUs in it, 28 threads per CPU, 56 threads total for system.

Are those a lot cheaper? Sounds like a good way to get the latest thing for less cash.

Engineering sample CPUs are cheaper but there are usually different versions of the samples. They don't have the same clocking specs as production CPUs (usually lower), there may be some bugs and they don't work in all motherboards. Research well before buying any.

Don't seem to be many on Ebay, but the one I saw was $300 instead of $1500!

I just realized why task info is reporting 2 CL devices. I also have a GTX 1050 Ti installed just to use for video output. The BOINC client is configured to ignore it but the milkyway still knows it's there. It doesn't use it but it sees it. I forget it's there.

For the engineering sample CPUs, search for Xeon V4 ES. Then you can filter the search results for the number of cores you want. Identify the QS code specification code like QH2N, QHV7, etc. Then research the ES spec on other forums and see if any is using it and anything they report about it, good or bad. I some 14 core QHV7s that haven't given any problems. Again, they don't work in all mobos. I use Supermicro X10DAL-i mobos on a few systems. Look for a seller that lists the compatible mobos

I just realized why task info is reporting 2 CL devices. I also have a GTX 1050 Ti installed just to use for video output. The BOINC client is configured to ignore it but the milkyway still knows it's there. It doesn't use it but it sees it. I forget it's there.

Handy for getting 600 instead of 300 WUs at a time.

For the engineering sample CPUs, search for Xeon V4 ES. Then you can filter the search results for the number of cores you want. Identify the QS code specification code like QH2N, QHV7, etc. Then research the ES spec on other forums and see if any is using it and anything they report about it, good or bad. I some 14 core QHV7s that haven't given any problems. Again, they don't work in all mobos. I use Supermicro X10DAL-i mobos on a few systems. Look for a seller that lists the compatible mobos

Thanks for the info.

All Nvidia gpus report 4GB of memory in BOINC because BOINC only uses an outdated 32bit API. 32bits can only report 4GB of addressable memory even when the gpu has much more memory installed.
This is just a reporting issue. All of the memory of a Nvidia gpu gets used. The BOINC developers have never acted on the reported issue or merged the provided solution that GPUUG Team came up with.
https://github.com/BOINC/boinc/issues/1773
You can modify the necessary code module and compile BOINC yourself to fix the issue and have your Nvidia cards report the proper amount of memory they contain.

---