I'm looking for a confirmation that the app works on windows machine with > 1 device. I'm seeing some

7:33:28 (10748): wrapper: running acemd3.exe (--boinc input --device 2)
# Engine failed: Illegal value for DeviceIndex: 2

I'm currently running test340-TONI_GSNTEST3-3-100-RND9632_0 on a GTX 1660 SUPER under Windows 7, BOINC v7.16.3

The machine has a secondary GPU, but is running on the primary: command line looks correct, as

"acemd3.exe" --boinc input --device 0

Progress is displaying plausibly as 50.000% after 2 hours 22 minutes, updating in 1% increments only.

Task completed and validated.

The %Progress keeps resetting to zero on 2080 Ti's but seems normal on 1080 Ti's.

My impression so far is that Win7-64 can run four WUs on two 1080 Ti's fine on the same computer fine.
The problem seems to be with 2080 Ti's running on Win7-64. I'm running four WUs on one 2080 Ti with four Einstein or four Milkyway on the second 2080 Ti seems ok so far. Earlier when I had two WUs on each 2080 Ti along with either two Einstein or two Milkyway that it kept resetting.
All Linux computers with 1080 Ti's seem normal.
Plan to move my two 2080 Ti's back to a Linux computer and try that.

---

The %Progress keeps resetting to zero on 2080 Ti's but seems normal on 1080 Ti's.

My impression so far is that Win7-64 can run four WUs on two 1080 Ti's fine on the same computer fine.
The problem seems to be with 2080 Ti's running on Win7-64. I'm running four WUs on one 2080 Ti with four Einstein or four Milkyway on the second 2080 Ti seems ok so far. Earlier when I had two WUs on each 2080 Ti along with either two Einstein or two Milkyway that it kept resetting.
All Linux computers with 1080 Ti's seem normal.
Plan to move my two 2080 Ti's back to a Linux computer and try that.

As a single ACEMD3 task can push the GPU to 100%, it would be interesting to see if there is any clear advantage to running multiple ACEMD3 tasks on a GPU.

@rod4x4 which error? no resume on different cards is known, please see the faq.

Hi Toni
Not referring to any particular error.
When the ACEMD3 task (Child task) experiences an error, the Wrapper always reports a generic error (195) in the Exit Status:
Exit status 195 (0xc3) EXIT_CHILD_FAILED

Can the specific (Child) task error be passed to the Exit Status?

Okay, my 1060 with Max-Q design completed one task and validated.

Okay, my 1060 with Max-Q design completed one task and validated.

Good news.
Did you make any changes to the config after the first failure?

My windows 10 computer on the RTX 2080 ti is finishing these WUs in about 6100 seconds, which is about the same time as computers running linux with same card.

Is the WDDM lag gone or is it my imagination?


The 18000 to 19000 seconds range are the these WUs running on the GTX 980 ti.



http://www.gpugrid.net/results.php?hostid=263612&offset=0&show_names=0&state=0&appid=32

@Rod 4*4. I did make a change but I do not know it's relevance. I set SWAN_SYNC to 0. I did that for some other reason. Anyway, second WU completed and validated.

Is the WDDM lag gone or is it my imagination?

Given that the various tool now show a GPU utilization of mostly up to 99% or even 100% (as it was with WinXP before), it would seem to me that the WDDM does not play a role any more.

WU now require 1 CPU core - WU run slower on 4/5 GPUs with (4) CPU cores.

3 GPUs full speed while 4 or 5 cause GPU usage to tank.

ELISA WU GPU max power draw (55C GPU temp)

330W on 2080ti 95% GPU utilization (PCIe 3.0 x4) @ 1995MHz

115W on 1660 99% "" (PCIe 3.0 x4) @ 1995MHz

215W on 2080 89% "" (PCIe 2.0 x1) @ 1995MHz

Progress bar runtime (2080ti 1hr 40min) / (2080 2hr 40min) / (1660 5hr)

1660 runtime equal to the 980ti.

My windows 10 computer on the RTX 2080 ti is finishing these WUs in about 6100 seconds, which is about the same time as computers running linux with same card.

Is the WDDM lag gone or is it my imagination?

I came to this conclusion too. The runtimes on Windows 10 are about 10880 sec (3h 1m 20s) (11200 sec on my other host), while on Linux it's about 10280 sec (2h 51m 20s) on GTX 1080 Ti (Linux is about 5.5% faster). These are different cards, and the fastest GPU appears to be the slowest in this list. It's possible that the CPU feeding the GPU(s) is more important for the ACEMD3 than it was for the ACEMD2, as my ACEMD3-wise slowest host has the oldest CPU (i7-4930k, which is 3rd gen.: Ivy Bridge E) while the other has an i3-4330 (which is 4rd gen.: Haswell). The other difference between the two Windows host is that the i7 had 2 rosetta@home tasks running, while the i3 had only the ACEMD3 running. Now I reduced the number of rosetta@home tasks to 1. I will suspend rosetta@home if there will be a steady flow of GPUGrid workunits.

Is the WDDM lag gone or is it my imagination?

Given that the various tool now show a GPU utilization of mostly up to 99% or even 100% (as it was with WinXP before), it would seem to me that the WDDM does not play a role any more.

While this high readout of GPU usage could be misleading, I think it's true this time. I expected this to happen on Windows 10 v1703, but apparently it didn't. So it seems that older CUDA versions (8.0) don't have their appropriate drivers to get around WDDM, but CUDA 10 has it.
I mentioned it at the end of a post almost 2 years ago.
There are new abbreviations from Microsoft to memorize (the links lead to TLDR pages, so click on them at your own risk):
DCH: Declarative Componentized Hardware supported apps
UWP: Universal Windows Platform
WDF: Windows Driver Frameworks
- KMDF: Kernel-Mode Driver Framework
- UMDF: User-Mode Driver Framework
This 'new' Windows Driver Framework is responsible for the 'lack of WDDM' and its overhead. Good work!

Hi,
many thanks for add new tasks, but please add again more tasks lol
bests regards
laurent from Belgium

100% GPU use and low WDDM overhead are nice news. However, they may be a specific to this particular WU type - we'll see in the future. (The swan sync variable is ignored and plays no role.)

For me, 100% on GPU is not the best ;-)
Because I have just one card on the pc, and I can't see videos when GPUgrid is running. Even if I ask to smplayer or vlc to use the CPU So I have to pause this project when I use my pc.
Maybe one day we will can put some priority to the use of GPU (on linux).
I think I will buy a cheap card for manage the TV and play movies. But well, in general I am at work or somewhere else...

Nice to have some work. Folding@Home will wait. I was thinking to change, the others BOINC projects running on GPU doesn't interest me.

there was a task which ended after 41 seconds with:
195 (0xc3) EXIT_CHILD_FAILED

stderr here: http://www.gpugrid.net/result.php?resultid=21514460

Are tasks being sent out for CUDA80 plan_class? I have only received new tasks on my 1080Ti with driver 418 and none on another system with 10/1070Ti with driver 396, which doesn't support CUDA100

there was a task which ended after 41 seconds with:
195 (0xc3) EXIT_CHILD_FAILED

stderr here: http://www.gpugrid.net/result.php?resultid=21514460

unfortunately ACEMD3 no longer tells you the real error. The wrapper provides a meaningless generic message. (error 195)
The task error in your STDerr Output is

# Engine failed: Particle coordinate is nan

I had this twice on one host. Not sure if I am completely correct as ACEMD3 is a new beast we have to learn and tame, but in my case I reduced the Overclocking and it seemed to fix the issue, though that could just be a coincidence.

For me, 100% on GPU is not the best ;-)
Because I have just one card on the pc, and I can't see videos when GPUgrid is running. Even if I ask to smplayer or vlc to use the CPU So I have to pause this project when I use my pc.
Maybe one day we will can put some priority to the use of GPU (on linux).
I think I will buy a cheap card for manage the TV and play movies. But well, in general I am at work or somewhere else...

Nice to have some work. Folding@Home will wait. I was thinking to change, the others BOINC projects running on GPU doesn't interest me.

C'est exactement ce que j'ai fait en installant une GT710 juste pour la sortie vidéo, c'est au top, du coup ma 980 Ti à 100% de charge ne me dérange pas du tout !

---