It seems to be working. At least several previously-failing hosts switched to success.

Also possibly working (please check)

* progress bar
* pause/resume

Will be un-marking beta soon (so more hosts get it).

v2.03 works on both my computers!

What was the change between v2.02 and v2.03 if you don't mind me asking. :)

The progress bar works.
I didn't try pause/resume.

It was a cryptic bug in the order loading shared libraries, or something like that. Otherwise unexplainably system-dependent.

I see VERY few failures now. The new app will be a huge step forward on several aspects, not least maintainability. We'll be transitioning gradually.

My host that failed on 2.02 now works with 2.03.

Great work

App 2.04 should support cuda10, if the scheduler collaborates.
Expect hiccups...

T

Well done Toni!

Much appreciated!

Toni Tests looking good on my machines, even on the 2080Ti!

Two hosts have received 2.04 and processed on the 2080 along with a 1070 successfully.

The other host is still on 2.03 I assume because I had the gpu_exclude running on it for the 2080 card. Removing the gpu_exclude for all hosts now.

It looks like we have a winner.

App 2.04 should support cuda10, if the scheduler collaborates.

Would it help the performance of a GTX 1070 (Ubuntu 16.04) to upgrade from CUDA 9.2 to CUDA 10 drivers?

From what we've seen over at Seti with our special app, the CUDA9 application running the 410 series driver is fastest for Pascal cards. Of course if you have a Turing card, you are forced into using the 418 series drivers with CUDA10.

Not sure if the same observation applies here at GPUGrid with its apps.

My guess and suggestion would be to just stand pat until somebody runs both the 2.03 and 2.04 apps against the same tasks in the benchmark utilities and proves otherwise.

I assume all the kit 'n kaboodle of files needed for the wrapper would have to be put into the benchmark tool. Not as simple as testing on Seti tasks.

Yes, it is a bit premature. I will await for the final app. Thanks for the input.

Yes, it is a bit premature. I will await for the final app. Thanks for the input.

Generally for crunching, a newer app is never necessary. The normal reason for updated drivers is to add ever more compatibility with newer games. And the drivers get ever more bloated and slower.

Only if new architecture arrives requiring newer drivers that can drive them is it really necessary to update drivers.

The low level math functions of the drivers have been pretty much static for years.

Of course the silicon gets ever faster which is where the most improvement occurs.

On one host with both a 2080 and 1070, the 2080 ran the test WU in 125 seconds versus 230 seconds with the 1070. Both on the 2.04 app.

On one host with both a 2080 and 1070, the 2080 ran the test WU in 125 seconds versus 230 seconds with the 1070. Both on the 2.04 app.

I see 93 seconds on my RTX 2080 Ti (v2.04)
(The v2.03 had failed before on my host btw)

I might be a little bit confused: Any time line when real production task of the 2.04 app will be supplied to Linux hosts? In the next days, weeks, months? I changed my BOINC settings to accommodate and give GPUGRID tasks the preference, however if there are no tasks available in the near future, I would like to change the settings again to better serve other projects.

Toni please answer as soon and as specific as possible. Thanks!

We'll be transitioning gradually.

Was Toni's last response. I would not expect the project to be as responsive as you request.
My suggestion is to move on to your other projects and occasionally check back in here for any news of non-beta apps and newly generated work units.

To summarize: acemd3 cuda80 and cuda100 are very satisfactory for linux. Very few failures, fast, blazing on rtx 2080. The win app will need work though.

There is definitely NO NEED TO install the CUDA toolkit for GPUGRID. It's not required, so don't complicate your set-up unnecessarily.

You just will want to have not-too-old drivers, however.

Toni, did you discover a cutoff limit for driver version you should not go below or you will increase your errors?

The most recent driver in the linux mint 19.1 repository is 390.116 (cuda 9).

You can get 4xx.xx proprietary drivers that support cuda 10 at this site by adding the PPA to your system. https://launchpad.net/~graphics-drivers/+archive/ubuntu/ppa

Or you can download and install the drivers from nvidia directly. The latter is a little more adventurous for linux users.

I was mainly asking if it was necessary to run the latest drivers like the 430 series. You have to run at least the 418 series to be compatible with Turing.

I have used the graphics-drivers ppa for years now. I always recommend that method for installing Nvidia drivers to Linux noobs as easiest.

The most recent driver in the linux mint 19.1 repository is 390.116 (cuda 9).

My Linux Mint 19.1 computers have all been using CUDA 10 for a long time and were installed using the Driver Manager:

aurum@Rig-27:~$ nvidia-smi
Sat Jun 8 14:16:29 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 430.14 Driver Version: 430.14 CUDA Version: 10.2

Side note, I know just enough Linux to be dangerous. Someone once told me to install the CUDA Toolkit to get CUDA 10 and it caused problems with Linux Mint 19. This is anecdotal but I think my computers generally work better without CUDA Toolkit installed.

---