I noticed today that I've had two separate replication _3 long tasks issued.

e6s17_e5s7p0f556-GERARD_CXCL12CHALCLD_mol39_1-0-1-RND3782
e26s8_e16s16p0f492-GERARD_CXCL12CHALCYDIM_chalcone266_1-0-1-RND4729

Five of the six previous attempts failed because they were sent to incompatible Pascal-class GPUs. But the system couldn't do anything about that without sending out all the files, waiting for the host to attempt processing them, and receiving an error message back. That's a heck of a waste of time and resources, for both the project and the volunteer.

These cards have been on public sale for a while now, and are getting quite common. If the software really hasn't caught up with the hardware yet, perhaps you could consider excluding Pascal GPUs from work allocation for the time being? The plan_class_spec.xml file on the server allows for

Fields for NVIDIA GPU apps

<max_nvidia_compcap>MMmm</max_nvidia_compcap>
maximum compute capability

which would do the trick.

... perhaps you could consider excluding Pascal GPUs from work allocation for the time being?

+1000

I am getting some work units but they error out immediately:

<core_client_version>7.6.22</core_client_version>
<![CDATA[
<message>
(unknown error) - exit code -59 (0xffffffc5)
</message>
<stderr_txt>
# GPU [GeForce GTX 1070] Platform [Windows] Rev [3212] VERSION [65]
# SWAN Device 0	:
#	Name		: GeForce GTX 1070
#	ECC		: Disabled
#	Global mem	: 4095MB
#	Capability	: 6.1
#	PCI ID		: 0000:01:00.0
#	Device clock	: 1683MHz
#	Memory clock	: 4004MHz
#	Memory width	: 256bit
#	Driver version	: r370_00 : 37306
#SWAN: FATAL: cannot find image for module [.nonbonded.cu.] for device version 610

</stderr_txt>
]]>

I am getting some work units but they error out immediately:

It's because the GPUGrid app does not support Pascal GPUs yet.

I'm curious why F@H supports Pascal GPUs just fine but GPUGRID doesn't?

Pascal also works just fine at every other DC project that I know, with CUDA 8.0 driver. However, GPU-Grid targets CUDA 8.5 for the next app and is waiting for its release. We don't know what CUDA 8.5 offers them beyond what CUDA 8.0 already provides.

MrS

---

Scanning for our furry friends since Jan 2002

FAH uses OpenCL, not CUDA, so the same WUs can run on any card.

But there's no answer here why 8.5 is needed vs another version from what I've seen.

But there's no answer here why 8.5 is needed vs another version from what I've seen.

I think MJH said (or implied) earlier that the bug they needed fixed was not fixed in CUDA 8.0, as they had originally hoped, but won't be implemented until CUDA 8.5. There seems to be no speculation on when that will be released, so Nvidia must keep a tight leash on their developers.

Poem@home being stopped, I think GPUGRID will get much more GPU power if available for Pascal.


http://boincstats.com/en/stats/-5/user/detail/839477/projectList

I've recognized too, they stopped some weeks ago. I had been running Poem on my 1070 for the last month, waiting for GPU grid to support Pascal. Now, it's doing Seti.
Anyone can give an outlook for Pascal support at GPU grid ?

There are speculations (on other projects [not BOINC, but development on projects like Debian/Ubuntu releases and such]) that the new branch of GCC was supposed to be supported in CUDA 8 and then it was put on speculation as Q1 2017. Could that be the answer we are looking for and can't get? Are we waiting on the GCC6 support with CUDA 8+? If so, the Q1 2017 speculation and a release of CUDA is speculated to follow that in short order just like the original GCC6 had the CUDA 8 release within about 20 days after realizing it was not going to have that support. Does BOINC, or some other component of what we are doing, like the MD/ACEMD rely on GCC support and to upgrade to the new architecture of NVIDIA GPUs in a "perfect world moving forward release" capacity, we need a CUDA 8 (or better) with GCC6 (or better) support?

I only thought of this after re-reading the threads here on this and came across this one https://www.gpugrid.net/forum_thread.php?id=4378&nowrap=true#44483 that I read, but missed the one part I just caught about GCC 5.4 breaking CUDA 8 RC. I think I missed it because I didn't understand it. I still can't say I do, but after searching around, I think I do a little.

There are 2 aspects of this questioning.
1) It is speculation on timing of CUDA 8.5 and GCC6 being in the first quarter of 2017. This may give us a time to expect this since there seems to be a blackout of information from NVDIA and those inside the CUDA toolkit developers teams on projects such as ours on when that release will come, so we must rely on the speculation of some of those developers.
2) It is a question, not a speculation on my part. Meaning that I am questioning the need of GCC6 compiling to be some part of what we need in a CUDA 8+ release. And if the answer to all of this that I wrote is simply, "No, that is not a factor", then the whole thing I just wrote is useless to read in conjunction with the rest of this thread. But if the answer is (or could be), "Yes. We do/might need that support in a CUDA release", then the speculation of the timeline of Q1 2017 for what we need is more clear and we can at least say that it is speculated to be around that time when people continue to come asking that "when" and "why" questions. That would also mean that the onus is not on GPUGRID (which we already kenw), but it is also not on NVIDIA (which we have been telling people), but on GNU to release the GCC6 version in order for NVIDIA to release a CUDA library with that GCC compiler compatibility to be compatible with the new applications compiled for the new hardware architecture.

P.S. I hope all this was clear to the reader as it was tough on me as the writer to keep/make it clear. I am on pain meds for kidney stones right now and keeping/making clear thoughts is difficult.) If it was not clear, please ask for clarification and I can re-read what I wrote and try to make it more clear to what I mean. Thanks. - Mike

Hello, Mike:

First, I hope you soon recover from your painful affliction.

I do not pretend to understand much of the material in this thread and have this question:

Will I be able to continue contributing to GPUGrid research with these two rigs:

1. AMD FX-8350 with two GTX 660 Ti cards,

2. AMD 1090T with two GTX-650 Ti cards

Thanks,

John

There are speculations (on other projects [not BOINC, but development on projects like Debian/Ubuntu releases and such]) that the new branch of GCC was supposed to be supported in CUDA 8 and then it was put on speculation as Q1 2017. Could that be the answer we are looking for and can't get? Are we waiting on the GCC6 support with CUDA 8+? If so, the Q1 2017 speculation and a release of CUDA is speculated to follow that in short order just like the original GCC6 had the CUDA 8 release within about 20 days after realizing it was not going to have that support. Does BOINC, or some other component of what we are doing, like the MD/ACEMD rely on GCC support and to upgrade to the new architecture of NVIDIA GPUs in a "perfect world moving forward release" capacity, we need a CUDA 8 (or better) with GCC6 (or better) support?

I only thought of this after re-reading the threads here on this and came across this one https://www.gpugrid.net/forum_thread.php?id=4378&nowrap=true#44483 that I read, but missed the one part I just caught about GCC 5.4 breaking CUDA 8 RC. I think I missed it because I didn't understand it. I still can't say I do, but after searching around, I think I do a little.

There are 2 aspects of this questioning.
1) It is speculation on timing of CUDA 8.5 and GCC6 being in the first quarter of 2017. This may give us a time to expect this since there seems to be a blackout of information from NVDIA and those inside the CUDA toolkit developers teams on projects such as ours on when that release will come, so we must rely on the speculation of some of those developers.
2) It is a question, not a speculation on my part. Meaning that I am questioning the need of GCC6 compiling to be some part of what we need in a CUDA 8+ release. And if the answer to all of this that I wrote is simply, "No, that is not a factor", then the whole thing I just wrote is useless to read in conjunction with the rest of this thread. But if the answer is (or could be), "Yes. We do/might need that support in a CUDA release", then the speculation of the timeline of Q1 2017 for what we need is more clear and we can at least say that it is speculated to be around that time when people continue to come asking that "when" and "why" questions. That would also mean that the onus is not on GPUGRID (which we already kenw), but it is also not on NVIDIA (which we have been telling people), but on GNU to release the GCC6 version in order for NVIDIA to release a CUDA library with that GCC compiler compatibility to be compatible with the new applications compiled for the new hardware architecture.

P.S. I hope all this was clear to the reader as it was tough on me as the writer to keep/make it clear. I am on pain meds for kidney stones right now and keeping/making clear thoughts is difficult.) If it was not clear, please ask for clarification and I can re-read what I wrote and try to make it more clear to what I mean. Thanks. - Mike

John, this is surely not the place for the question, but since you asked, I'll answer.

If you look on this page https://www.gpugrid.net/performance.php at the bottom, you will see both of your cards listed in the currently "top ranked" cards while even some of mine doing the work quite well are not. The one has a listed 'longest run' of about 65 hours (about 2.75 days) and the timeout for work units is 5 days (120 hours). Short runs take much less time and are currently available, though that has not been the usual over the past year. Availability may change and the length of time needed to complete a task might change, but I think currently and maybe for quite a while, those cards will do the tasks in more than enough time. The nuances will happen depending on your mother board, PCI Express slot speeds, CPU availability, OS, and other tweaks and things that can be done and are listed on other threads.

The new cards are offering more opportunity, not a replacement of older opportunity, although with time, some architectures do become "extinct" after many years. As you can see (not clearly, but its there) on this page https://www.gpugrid.net/graphs/ that your level of Kepler cards are still doing most of the work for GPUGRID currently and Fermi cards are falling off with their being "taken off the market" in new code and general usefulness to complete work on time.

Crunch away sir!

we will have support for Pascal very soon, in a matter of days.

we will have support for Pascal very soon, in a matter of days.

This is good news.
Could you share more details about the new app?
What CUDA / driver version will be needed for it?
The old apps (CUDA60 & CUDA65) will still be available after the new app is released?
Will it work on older cards?
Will it work on older OS? (especially Windows XP - but I guess it depends on the driver)
For what OSes will it be available? (I guess it will come out for Linux first)

we will have support for Pascal very soon, in a matter of days.

Very good news! Time to start reducing my cache of Einstein tasks.

we will have support for Pascal very soon, in a matter of days.

Sweet! Now is finally the time to start looking into the cards available to upgrade.... but need to get my pcs stable!!! (my seeming constant struggle I keep losing)

we will have support for Pascal very soon, in a matter of days.

Sweet! Now is finally the time to start looking into the cards available to upgrade.... but need to get my pcs stable!!! (my seeming constant struggle I keep losing)

What's wrong with the PC?

---

Cruncher/Learner in progress.

we will have support for Pascal very soon, in a matter of days.

Any recent change in the server scheduling for Pascal (also with Maxwell) blacklisting?
Unable to download tasks during last couple days on my (2) GTX 1070 / (1) GTX 1060 / (2) GTX 970 / (1) 384core GT 630 system.
Windows 8.1 refuses to see more than 4 GPU's since r370 branch (including r375). As of late a couple of 7/8 GPUs Window 10 (enterprise) were on the performance page.

What are we waiting for at this point? The developer of your software to update to the new Cuda?