If the zi3v code was originally targeted at CUDA8, I think we should try that first. Should only need the download and installation of the CUDA8 package and then change the configuration to point at the CUDA8 directories. I need to speak with Sam about that. We have given up on the 0.97 code branch.

. . 0.97 works fine for me :) It is a pity it doesn't for CUDA60 ...

Stephen

? ?

Stephen, I am confusing you obviously with carrying on two conversations in the thread. There is no reason why the zi3v app couldn't be compiled for x86_64 with the CUDA 8 toolkit for the GT 730 in Linux x86_64.

https://developer.nvidia.com/cuda-80-ga2-download-archive

I am just screwed that I don't have any choice for a lesser CUDA toolkit other than CUDA10 for my Jetson Nano Tegra X1 ARM64 platform.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Actually, it may work with 6.5. It says *aarch64 - 64-bit ARM CPU architecture, like that found on Jetson TX1 onwards* https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#cross-platform You just need to install Ubuntu 14.04, read the manual, and chose the correct packages.
Good Luck, and Have Fun ;-)

Thanks for that TBar. I actually wasn't thinking of cross-compiling on Ubuntu for aarch64. I was just thinking of compiling natively like we have done so far on the Nano.

I think cross-compiling in Ubuntu 14.04 might be the only way to accomplish getting the zi3v app compiled for CUDA6.5.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

After looking at it a little harder, it seems the only ARMv8 64-bit Toolkit is the one meant to run on the aarch64 device itself, and is labeled "Generic CUDA Toolkit", https://developer.nvidia.com/cuda-toolkit-65
The one for Ubuntu 14.04 Cross-Compile is only 32 bit. So, try removing the cuda 10 Toolkit and see if the one here works on the Tegra X1, http://developer.download.nvidia.com/compute/cuda/6_5/rel/installers/cuda_6.5.14_linux_aarch64_native.run It does appear 6.5 is the only Toolkit with a section labeled Linux ARM...

Stephen, I am confusing you obviously with carrying on two conversations in the thread. There is no reason why the zi3v app couldn't be compiled for x86_64 with the CUDA 8 toolkit for the GT 730 in Linux x86_64.
https://developer.nvidia.com/cuda-80-ga2-download-archive
I am just screwed that I don't have any choice for a lesser CUDA toolkit other than CUDA10 for my Jetson Nano Tegra X1 ARM64 platform.

. . OK, well TBar must not have used that resource or option when he compiled the CUDA80 version he made available because when I tested for him, while it ran on the GT730 it produced a lot of results with errors, though not all were in error. But His Cuda60 release worked without errors. I cannot follow up on that because my GT730 is now in a new home ...

Stephen

< shrug >

Well the ball is again in my court. Just need to install the CUDA6.5 toolkit and make another run at compiling. I really want to drop the number of Inconclusives even if the crunch times go longer.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Two quick questions about upgrading from the special app that runs CUDA 90 to the one that runs CUDA101:

1. Will the 101 WUs crunch faster than the 90WUs?
2. If so, do I have to exhaust my cache before the upgrade to avoid computation errors?

The hosts I am looking to upgrade:

https://setiathome.berkeley.edu/show_host_detail.php?hostid=8657594

https://setiathome.berkeley.edu/show_host_detail.php?hostid=8706475

I would say to not bother since the CUDA101 is more suited for the new Turing cards to see any performance benefit. Since you have older Pascal low-mid range cards, I would just stick with the CUDA90 app.

If you do decide to change don't do anything more than change the appname in the two places in the app_info. Don't do anything else. You don't have to run your cache down at all. Just stop BOINC, edit the new filename, save the app_info and restart BOINC

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

. . @ Keith

. . Speaking of updating, as a volunteer who has used both Cuda90 0.97 and 0.98b over several platforms. would you say there is any significant improvement in run times with the latter version? That is, do you feel there has been any improvement in efficiency?

Stephen

? ?

I would say there is an improvement in the 0.98 code compared to 0.97. I proved that to myself when I benched the betas for Petri before wide release. But Ian states that his beta of the 0.98b1 CUDA10 with driver version 410 beats out the 0.98b1 CUDA101 version with drivers 418. He kept his beta release and is running that to great effect. I was not a tester of that version so never got a chance to benchmark it.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I would say to not bother since the CUDA101 is more suited for the new Turing cards to see any performance benefit. Since you have older Pascal low-mid range cards, I would just stick with the CUDA90 app.

If you do decide to change don't do anything more than change the appname in the two places in the app_info. Don't do anything else. You don't have to run your cache down at all. Just stop BOINC, edit the new filename, save the app_info and restart BOINC

Thanks for the info. Not knowing that the 101 is for Turing cards I am running the 101 on two hosts with GTX 750 Ti cards. Would it make sense to downgrade to 90 or just leave well enough alone?

Don't know for the low level cards how they would respond to the CUDA90 app. Try it and report back.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Don't know for the low level cards how they would respond to the CUDA90 app. Try it and report back.

I am already running the CUDA90 app on similar cards and it seems pretty fast to me. See this host
https://setiathome.berkeley.edu/show_host_detail.php?hostid=8657594

Try to find a host with a 1060 that is running the CUDA101 app for comparison.

[Edit] Or download Rick's benchmark tool and do the comparison between the two apps yourself in a controlled test with the same WU's.

https://github.com/Ricks-Lab/benchMT

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Try to find a host with a 1060 that is running the CUDA101 app for comparison.

Like this 1?

https://setiathome.berkeley.edu/show_host_detail.php?hostid=8150019

Cheers.

Comparing the two looks like the CUDA101 app has the best and most consistent times. Loren, you have more cpu resources and I don't see any cpu tasks being run or at least not for Seti, so lack of cpu support for the gpu tasks is not likely. The 1060 3GB cards on Vyper's host seem to be turning in better times and more consistently.

I would say go for the CUDA101 app and see if you can replicate the same times as Vyper's host.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Just replacing those two lines in the app_info.xml caused computation errors. So I am setting NNT and will then change to CUDA101.

So, are you running the required 418 series drivers? The readme states that is necessary to run the CUDA101 application in the same sentence it tells you to swap appnames in the app_info file.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Yeah I’m guessing he tried it on the system with 396 drivers.

He currently has 2 qualifying systems. One running 418 drivers and another running 430.

He has a 3rd system with a 1050ti and 390 drivers, and a 4th with some 1060s and 396 drivers. It looks like that’s the one he tried. He’ll have to upgrade the drivers to run 10.1 on those systems.

---

Seti@Home classic workunits: 29,492 CPU time: 134,419 hours

I would say there is an improvement in the 0.98 code compared to 0.97. I proved that to myself when I benched the betas for Petri before wide release. But Ian states that his beta of the 0.98b1 CUDA10 with driver version 410 beats out the 0.98b1 CUDA101 version with drivers 418. He kept his beta release and is running that to great effect. I was not a tester of that version so never got a chance to benchmark it.

. . Would you say that improvement (0.98b over 0.97) is enough to bother updating? I am running 0.97 on both the rig with 970s and the rig with the 1050ti (which sadly is using the repository version of BOINC) so after reading things like the comment that the 101 version is only useful on Turing cards I am not confident it would make any significant difference on my rigs.

Stephen

? ?