Advanced search

Message boards : Graphics cards (GPUs) : Turing support status at end of thread (former RTX 2080 TI thread)

Author Message
BJ
Send message
Joined: 23 Aug 09
Posts: 1
Credit: 289,044,901
RAC: 78,731
Level
Asn
Scientific publications
watwatwatwatwat
Message 50289 - Posted: 25 Aug 2018 | 20:32:50 UTC

Hi,

Is the new RTX 2080 or 2080 Ti supported?

Thanks,

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50290 - Posted: 25 Aug 2018 | 22:33:29 UTC - in response to Message 50289.

Is the new RTX 2080 or 2080 Ti supported?

Not yet.

jjch
Send message
Joined: 10 Nov 13
Posts: 28
Credit: 12,212,147,939
RAC: 5,575,659
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 50303 - Posted: 27 Aug 2018 | 17:47:29 UTC

I think I will be better off sticking with the GTX 1080's for awhile anyway. The prices are still reasonable and the power consumption is lower for a decent amount of compute power.

It will really be interesting to see the actual compute power benchmarks for the 20 Series before even considering blowing that kind of money on one. Going to need systems with bigger power supplies for anything more than a 2070 too.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 50517 - Posted: 15 Sep 2018 | 2:20:30 UTC

Since the architecture embargo lifted today along with unboxing(who cares)I find it interesting about the differences in the SM architectures between Pascal and Turing. I think we might be pleasantly surprised by compute improvements.Turing shares more commonalities with Volta than it does with Pascal.

What I think may make the biggest differences in compute performance may come down to simultaneous FP64 and INT32 instructions running through the SM without penalty. Also the larger shared memory instructions caches will likely be a major boost. And finally, the bandwidth improvement simply from using the higher clocked GDDR6 over the GDDR5/X memory should help.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 50564 - Posted: 20 Sep 2018 | 0:07:20 UTC

Compute benchmarks are out today at Anandtech. 200% faster on Geekbench 4 OpenCL compute benchmarks than Pascal. 30-40% faster on Folding@Home benchmarks.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 50591 - Posted: 21 Sep 2018 | 23:11:28 UTC

CUDA/OpenCL compute benchmarks out at Phoronix.com.

NVIDIA GeForce RTX 2080 Ti Shows Very Strong Compute Performance Potential

Profile Chilean
Avatar
Send message
Joined: 8 Oct 12
Posts: 97
Credit: 385,268,298
RAC: 17
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 50598 - Posted: 25 Sep 2018 | 12:36:39 UTC

Shouldn't higher Compute Cap be backwards compatible for code written for a lower CC?
____________

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 319,920
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50599 - Posted: 25 Sep 2018 | 20:34:22 UTC

Usually GPU-Grid is not compatible with new architecture generations for a few weeks. So I guess we're waiting for the first one to get a card, try and report to the team.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile Chilean
Avatar
Send message
Joined: 8 Oct 12
Posts: 97
Credit: 385,268,298
RAC: 17
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 50608 - Posted: 26 Sep 2018 | 15:04:52 UTC - in response to Message 50599.

Usually GPU-Grid is not compatible with new architecture generations for a few weeks. So I guess we're waiting for the first one to get a card, try and report to the team.

MrS


In this thread someone stated (along myself) that V100 weren't supported, failing with this message:

#SWAN: FATAL: cannot find image for module [.nonbonded.cu.] for device version 700


https://www.gpugrid.net/forum_thread.php?id=4352
____________

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 319,920
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50657 - Posted: 8 Oct 2018 | 19:53:06 UTC - in response to Message 50608.

Indeed (RTX2080):

SWAN: FATAL: cannot find image for module [.nonbonded.cu.] for device version 750

____________
Scanning for our furry friends since Jan 2002

Profile Chilean
Avatar
Send message
Joined: 8 Oct 12
Posts: 97
Credit: 385,268,298
RAC: 17
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 50658 - Posted: 8 Oct 2018 | 21:22:57 UTC
Last modified: 8 Oct 2018 | 21:23:12 UTC

Shouldn't this get fixed by simply allowing the app to understand what a 700 or 750 device is?
____________

WPrion
Send message
Joined: 30 Apr 13
Posts: 69
Credit: 1,000,072,969
RAC: 1
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 50660 - Posted: 9 Oct 2018 | 12:07:33 UTC - in response to Message 50599.

Usually GPU-Grid is not compatible with new architecture generations for a few weeks....


I'm confused about "compatibility". According to the performance page there are 32 RTX 2080's crunching away.

Win

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 508
Credit: 4,384,289,651
RAC: 1,564,248
Level
Arg
Scientific publications
watwatwat
Message 50661 - Posted: 9 Oct 2018 | 13:44:43 UTC

That "32" number is most likely the amount of WUs a 2080 has returned because if you look at the 1080ti, I don't think there are over 2000 1080tis crunching away here.
It looks like there are perhaps only 1-2 2080s here but as for why they work is actually a mystery. It took a long time to gain Pascal support in this application due to the new CUDA version. This is also a new CUDA version so I am not sure why it works.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50662 - Posted: 9 Oct 2018 | 15:05:59 UTC - in response to Message 50660.

I'm confused about "compatibility". According to the performance page there are 32 RTX 2080's crunching away.

Where does this 32 came from?
I can't find it.
The GPUs are listed on the bottom of the performance page, in the blue box plot.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50663 - Posted: 9 Oct 2018 | 15:12:01 UTC - in response to Message 50661.

It looks like there are perhaps only 1-2 2080s here but as for why they work is actually a mystery.

They don't work. If you click on the given result's id, you can see that it was crunched on a different GPU (usually on a GTX1080Ti), but the user changed their GPU in the meantime.

WPrion
Send message
Joined: 30 Apr 13
Posts: 69
Credit: 1,000,072,969
RAC: 1
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 50667 - Posted: 10 Oct 2018 | 11:52:05 UTC - in response to Message 50663.
Last modified: 10 Oct 2018 | 11:52:32 UTC

I just looked again and can't find the RTX data. I don't think I imagined it yesterday. Anyway, your answer makes sense. Thanks.

I wanted to understand if it was just the "tensor core" aspect of the RTX that was not supported yet, or the whole dang thing.

Win

Profile Chilean
Avatar
Send message
Joined: 8 Oct 12
Posts: 97
Credit: 385,268,298
RAC: 17
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 50680 - Posted: 12 Oct 2018 | 13:14:34 UTC - in response to Message 50667.

I just looked again and can't find the RTX data. I don't think I imagined it yesterday. Anyway, your answer makes sense. Thanks.

I wanted to understand if it was just the "tensor core" aspect of the RTX that was not supported yet, or the whole dang thing.

Win


Pretty sure none of the other projects that do run on 20XX use the tensor cores. I really don't understand why GPUGRID doesn't run on these. Higher CUDA version cards should be backward compatible with lower version CUDA code... (right?)
____________

mmonnin
Send message
Joined: 2 Jul 16
Posts: 265
Credit: 647,845,139
RAC: 60
Level
Lys
Scientific publications
wat
Message 50684 - Posted: 12 Oct 2018 | 20:54:02 UTC

Pascal took a while before there was a compatible app. Not as long as Asteroids but some update seems to be required. Newer cards can run older cuda apps for sure but it might be that older apps don't recognize newer cards as having cuda hardware.

rod4x4
Send message
Joined: 4 Aug 14
Posts: 105
Credit: 1,629,251,769
RAC: 1,191,273
Level
His
Scientific publications
watwatwatwatwatwatwat
Message 50837 - Posted: 7 Nov 2018 | 7:10:45 UTC

On the Performance Tab, the "GPU Performance Ranking" chart (Blue chart at bottom of page) currently shows 55 RTX2080ti GPUs.
Ave=6.65455h
Min=2.7h
Max=11.5h

It is the 10th GPU from the right hand side, in between TITAN X and Tesla M60.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 50838 - Posted: 7 Nov 2018 | 9:49:47 UTC - in response to Message 50837.

On the Performance Tab, the "GPU Performance Ranking" chart (Blue chart at bottom of page) currently shows 55 RTX2080ti GPUs.
Ave=6.65455h
Min=2.7h
Max=11.5h

It is the 10th GPU from the right hand side, in between TITAN X and Tesla M60.

This time I can see it too, but I'm sure it's a bug in the way the Performance tab collects the data for this statistics, as it's not prepared to handle hosts with multiple GPUs, and hosts with recently replaced (upgraded) GPUs.

rod4x4
Send message
Joined: 4 Aug 14
Posts: 105
Credit: 1,629,251,769
RAC: 1,191,273
Level
His
Scientific publications
watwatwatwatwatwatwat
Message 50839 - Posted: 7 Nov 2018 | 10:32:14 UTC - in response to Message 50838.
Last modified: 7 Nov 2018 | 11:30:25 UTC

This time I can see it too, but I'm sure it's a bug in the way the Performance tab collects the data for this statistics, as it's not prepared to handle hosts with multiple GPUs, and hosts with recently replaced (upgraded) GPUs.


If this is the case it would suggest the report is badly coded. If it does not take into account recently replaced GPUs or hosts with multiple GPUs, it is actually reporting on the host rather than the GPU.
Additionally, if card rotation is the cause, it would be reasonable to expect RTX2080 and RTX2070 GPUs listed most of the time as well (not just sometimes). Probability would have the card rotation happening randomly with these "phantom" cards always visible.
It is certainly perplexing whatever the case may be.

jjch
Send message
Joined: 10 Nov 13
Posts: 28
Credit: 12,212,147,939
RAC: 5,575,659
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 50844 - Posted: 8 Nov 2018 | 21:31:18 UTC

From what I have seen GPUGRID reports the greatest GPU installed in the host.
If there are two or more you could have both a RTX 2080 and a GTX 1080 in one host and it would report as the RTX 2080. This is true for the performance tab as well as the computers tab.

You could very easily be running tasks on a GTX 1080 and see that it is reporting as a RTX 2080 at the high level. If you drill down to the task level report you will see the actual card the task ran on.

rod4x4
Send message
Joined: 4 Aug 14
Posts: 105
Credit: 1,629,251,769
RAC: 1,191,273
Level
His
Scientific publications
watwatwatwatwatwatwat
Message 50845 - Posted: 9 Nov 2018 | 2:49:55 UTC - in response to Message 50844.

From what I have seen GPUGRID reports the greatest GPU installed in the host.

Yes. That certainly sound like what is happening with data captured and hence skewing results in this chart. (I would hope a RTX2080ti would get better results as reported by the GPU Ranking Performance chart)
It seems the GPU starting the task is not captured as part of the data set, just the host.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 50846 - Posted: 9 Nov 2018 | 3:55:26 UTC - in response to Message 50845.

Unfortunately, that is not the case. BOINC has a few foibles where it either can report the lowest PCIe BusID card in the host or the last card installed into the host and the drivers updated to pick it up.

This idiosyncrasy has been discussed ad nauseum in the Seti Number Crunching forum. The only way to tell what cards are actually installed in any host is to look at the stderr.txt output of a completed task.

It is not as simple to think the most powerful card is how the host's gpu capabilities are described.

mmonnin
Send message
Joined: 2 Jul 16
Posts: 265
Credit: 647,845,139
RAC: 60
Level
Lys
Scientific publications
wat
Message 50852 - Posted: 10 Nov 2018 | 2:23:59 UTC

My 1070 Ti/1070 is always shown as 2x 1070s since it is installed on top. Ti was added later.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 50853 - Posted: 10 Nov 2018 | 7:32:10 UTC - in response to Message 50852.

Then your 1070 likely has a lower BusID than the 1070Ti

Profile bcavnaugh
Send message
Joined: 8 Nov 13
Posts: 54
Credit: 830,150,450
RAC: 3,149,195
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwat
Message 50951 - Posted: 23 Nov 2018 | 2:13:44 UTC
Last modified: 23 Nov 2018 | 2:14:48 UTC

Any News for the RTX Cards?

Stderr output
<core_client_version>7.14.2</core_client_version>
<![CDATA[
<message>
(unknown error) - exit code -59 (0xffffffc5)</message>
<stderr_txt>
# GPU [GeForce RTX 2080 Ti] Platform [Windows] Rev [3212] VERSION [80]
# SWAN Device 0 :
# Name : GeForce RTX 2080 Ti
# ECC : Disabled
# Global mem : 11264MB
# Capability : 7.5
# PCI ID : 0000:01:00.0
# Device clock : 1635MHz
# Memory clock : 7000MHz
# Memory width : 352bit
# Driver version : r416_69 : 41694
#SWAN: FATAL: cannot find image for module [.nonbonded.cu.] for device version 750

</stderr_txt>
]]>
____________

Crunching@EVGA The Number One Team in the BOINC Community. Folding@EVGA The Number One Team in the Folding@Home Community.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 50983 - Posted: 30 Nov 2018 | 1:21:52 UTC

Without re-reading through the entire thread, has anyone running Linux and using a Turing card tried to see if it will run the Linux app?

I am getting a RTX 2080 card tomorrow and assume I will have to stop crunching for this project on that host until a new compatible app is released.

Or is all the comments about incompatibility been for Windows hosts?

Zalster
Avatar
Send message
Joined: 26 Feb 14
Posts: 177
Credit: 4,121,030,726
RAC: 52,749
Level
Arg
Scientific publications
watwatwat
Message 50984 - Posted: 30 Nov 2018 | 1:46:37 UTC - in response to Message 50983.

From what I remember no Linux or windows is working with the 20X0s cards.


____________

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 51007 - Posted: 6 Dec 2018 | 2:57:38 UTC

Just put my 2080 online with my projects under Linux. No bueno. It doesn't matter what OS the card runs under. No compatible app for Turing found. So I will have to remove that host from GPUGrid until an app for Turing shows up.

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 508
Credit: 4,384,289,651
RAC: 1,564,248
Level
Arg
Scientific publications
watwatwat
Message 51014 - Posted: 8 Dec 2018 | 21:35:51 UTC - in response to Message 51013.

I remember when gpu grid go from gtx580 to gtx680(perfect card) some "top" crunchers(zoltan,stoneage) have "indirectly" ask to scientifist here to little BRAKE in compiling to another generations of nvidia gf cards.. for 10 usd i will give you screnshoots where they indirectly asking in discuss here..)) same of these "top crunchers" people now watch with scary in eyes on AURUM-god becouse he will crush that dinosaurs on top 1, 2, ,3 ..
And of course performance on gpugrid with RTX20xx cards will add only if zoltan and "stoneFman" will report they buy ..)) now all are brake" and we sit here in old generations becouse ...
becouse university can not keep post graduant students.soo always we wait for new..and its same circle of "brake"

This is the most tinfoil hat comment I've ever read. I can assure you the top crunchers want nothing more than to be able to use the fastest, most power efficient cards and competition is merely some fun and incentive to buy more cards.

As for why they haven't upgraded the app. Priorities. Clearly their priorities are in the science itself at the moment and Matt Harvey, which was their app writer, has left. They are making by with what they have.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 51015 - Posted: 9 Dec 2018 | 1:11:47 UTC - in response to Message 51014.

Seems to be contagious. We haven't had any staff app developers at Seti for many years now. Now one of our last volunteer developers has called it quits too.

Unless the project volunteers donate a bunch of money to fund adding additional staff to a project, then what apps are currently available are the best we can expect.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51016 - Posted: 10 Dec 2018 | 0:11:20 UTC - in response to Message 51013.

I remember when gpu grid go from gtx580 to gtx680(perfect card) some "top" crunchers(zoltan,stoneage) have "indirectly" ask to scientifist here to little BRAKE in compiling to another generations of nvidia gf cards.. for 10 usd i will give you screnshoots where they indirectly asking in discuss here..))
I'll spare that 10 bucks...

same of these "top crunchers" people now watch with scary in eyes on AURUM-god becouse he will crush that dinosaurs on top 1, 2, ,3 ..
1. We're on each other's team.
2. Even if someone will overtake me on the toplist (BTW it happened before), he/she couldn't take away that 9 years while I was on the 2nd place. This is a very long time.
3. This "competition" of the volunteers is good for the project, because in reality this is cooperation.

And of course performance on gpugrid with RTX20xx cards will add only if zoltan and "stoneFman" will report they buy ..))
I buy GPUs for GPUGRID purposes only, and from the pace of the app development here I've learned that I should not buy new generation GPUs before the GPUGRID app supports it.
But I'll definitely buy a GTX 2080Ti if I can use it for GPUGRID.
I think that the GPUGRID staff is aware that the project is loosing computing power gradually, as the (non-GPUGRID enthusiast) volunteers are upgrading their GPUs. Believe me, that I (and AURUM too) can't replace a 100 other volunteers with RTX 2080 Tis.

now all are brake" and we sit here in old generations becouse ...
becouse university can not keep post graduant students.soo always we wait for new..and its same circle of "brake"
Then we should arrange a little fund raising campaign for the GPUGRID project to be able to hire a CUDA programmer who'll update the application.

Profile Chilean
Avatar
Send message
Joined: 8 Oct 12
Posts: 97
Credit: 385,268,298
RAC: 17
Level
Asp
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 51017 - Posted: 10 Dec 2018 | 13:15:03 UTC - in response to Message 51014.

This is the most tinfoil hat comment I've ever read.


Let me introduce you to Flat Earthers then lol


____________

JoergF
Avatar
Send message
Joined: 20 Apr 15
Posts: 283
Credit: 1,101,827,807
RAC: 111,337
Level
Met
Scientific publications
watwatwat
Message 51018 - Posted: 10 Dec 2018 | 13:36:00 UTC
Last modified: 10 Dec 2018 | 13:40:19 UTC

I've learned that I should not buy new generation GPUs before the GPUGRID app supports it.


I agree... and in addition to that, Turing has not improved much in regard to power efficiency compared to Pascal. Nvidia saved some GPU space by the shrink on one hand but used that saving for introduction of new Raytracing cores on the other hand. So for GPUGRID crunchers, there is hardly any reason to pension all the 1070 and 1080ies off.

same of these "top crunchers" people now watch with scary in eyes on AURUM-god becouse he will crush that dinosaurs on top 1, 2, ,3 ..


No need to go green with envy. The top ten RAC are impressive, but the contribution of the other crunchers should not be underestimated. There are still hundreds of GTX 750, 760, 960 or 970 which yield impressive numbers as well in their entirety. I would even assume that these are the backbone of GPUGRID (I think that is also what Zoltan suggested below), so I am pleased about having short runs again recently.

Then we should arrange a little fund raising campaign for the GPUGRID project to be able to hire a CUDA programmer who'll update the application.


Yes, but in many cases, money is not the only problem. It is generally difficult to even find someone who is willing to do the job. Information scientists are in great demand nowadays.
____________
I would love to see HCF1 protein folding and interaction simulations to help my little boy... someday.

tullio
Send message
Joined: 8 May 18
Posts: 157
Credit: 40,334,545
RAC: 9,490
Level
Val
Scientific publications
wat
Message 51037 - Posted: 13 Dec 2018 | 20:52:12 UTC

I see Tesla k40 and k80 cards advertised on theregister.co.uk for prices up to 3600 USD and they are not even new but "refurbished".
Tullio

Penguin
Send message
Joined: 14 Apr 18
Posts: 1
Credit: 24,779,512
RAC: 0
Level
Pro
Scientific publications
wat
Message 51083 - Posted: 25 Dec 2018 | 16:14:03 UTC - in response to Message 51018.
Last modified: 25 Dec 2018 | 16:14:31 UTC



Then we should arrange a little fund raising campaign for the GPUGRID project to be able to hire a CUDA programmer who'll update the application.


Yes, but in many cases, money is not the only problem. It is generally difficult to even find someone who is willing to do the job. Information scientists are in great demand nowadays.


I like the above idea. Shocking out of all the science people here, or maybe there's not as many as I thought, a programmer wouldn't step up to the challenge. I'm way out of programming experience haven't coded in 25 years, but surely some member hear would do it.



Any News for the RTX Cards?

Stderr output
<core_client_version>7.14.2</core_client_version>
<![CDATA[
<message>
(unknown error) - exit code -59 (0xffffffc5)</message>
<stderr_txt>
# GPU [GeForce RTX 2080 Ti] Platform [Windows] Rev [3212] VERSION [80]
# SWAN Device 0 :
# Name : GeForce RTX 2080 Ti
# ECC : Disabled
# Global mem : 11264MB
# Capability : 7.5
# PCI ID : 0000:01:00.0
# Device clock : 1635MHz
# Memory clock : 7000MHz
# Memory width : 352bit
# Driver version : r416_69 : 41694
#SWAN: FATAL: cannot find image for module [.nonbonded.cu.] for device version 750

</stderr_txt>
]]>


bcavnaugh good to see you here. And yeah, my 2080 is a no go here so far as well. Luckily i still have my 1070 as a 2nd card installed.

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 51086 - Posted: 25 Dec 2018 | 21:49:12 UTC

Anybody know where the science app repository is held? I would sure like to look at the nonbonded.cu module and the code which probes for the API and CC capability. I would think that a simple code change is all that is needed.

But you would also have to know what the compile flags are and what required library resources are needed to compile the app other than the obvious CUDA library.

Maybe after the holidays staff personnel will have some time to fix the app.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 319,920
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51089 - Posted: 25 Dec 2018 | 22:29:12 UTC

I don't think it's this easy. I'd be surprised if they couldn't fix the "if" clause themselves. However, releasing the app into the wild would require some serious testing, whether the results indeed match expectations and whether the same code path as Pascal works. Here we're not simply talking about runtime errors, but also more subtile issues like "slightly wrong" results. They probably have test routines and benchmark cases in place, but using and analyzing that is what requires the most time.

And the source code is not open. In fact, the core algorithmn is commercialized by a company GDF started a few years back, as far as I know. It's certainly not as simple as looking up some repository, at minimum an NDA would be required I guess.

MrS
____________
Scanning for our furry friends since Jan 2002

Aurum
Send message
Joined: 12 Jul 17
Posts: 118
Credit: 7,469,926,993
RAC: 3,552,085
Level
Tyr
Scientific publications
wat
Message 51093 - Posted: 26 Dec 2018 | 14:12:38 UTC - in response to Message 51014.
Last modified: 26 Dec 2018 | 14:13:52 UTC

...watch with scary in eyes on AURUM-god...


Bow to God Aurum lest he smite thee.

Aurum
Send message
Joined: 12 Jul 17
Posts: 118
Credit: 7,469,926,993
RAC: 3,552,085
Level
Tyr
Scientific publications
wat
Message 51094 - Posted: 26 Dec 2018 | 15:17:12 UTC - in response to Message 51016.
Last modified: 26 Dec 2018 | 15:20:30 UTC

2. Even if someone will overtake me on the toplist he couldn't take away that 9 years while I was on the 2nd place.
3. This "competition" of the volunteers is good for the project, because in reality this is cooperation.

Zoltan, I'm in awe of the fact that you recognized a worthy cause and have stuck to it for so long. I hope we can compete for many years to come. I read everything written in these forums and make notes. E.g., SWAN_SYNC seems to work well on my Win7 rigs. (I would post more but I have a very hard time communicating with this website from the States. It's taken over 50 tries to post this.)

I buy GPUs for GPUGRID purposes only, and from the pace of the app development here I've learned that I should not buy new generation GPUs before the GPUGRID app supports it.
But I'll definitely buy a GTX 2080Ti if I can use it for GPUGRID.
I think that the GPUGRID staff is aware that the project is loosing computing power gradually, as the (non-GPUGRID enthusiast) volunteers are upgrading their GPUs. Believe me, that I (and AURUM too) can't replace a 100 other volunteers with RTX 2080 Tis.

I'm a bit lost here, I think you're saying that if someone gets a 2080 Ti then they'll go elsewhere because GPUgrid doesn't talk to 2080s yet. But if they're a GPUgrid enthusiast they'd wait until 2080s can run GPUgrid WUs before getting one.
I also track the performance of GPUs as a way to judge their relative price. I got most of my GPUs on FleaBay at nice discounts. This is the data I track:
https://docs.google.com/spreadsheets/d/1v5gXral3BcFOoXs5n1M6l_Uo3pZpQYogn6gVlxRPnz0/edit#gid=0
https://docs.google.com/spreadsheets/d/1vcVoSVtamcoGj5sFfvKF_XlvuviWWveJIg_iZ8U2bf0/pub?output=html

Right now I'd say buy an EVGA 1070 Ti for under $425 (including shipping), an EVGA 1080 for $350, or an EVGA 1080 Ti for under $600 versus $1,000 for the bottom bin EVGA 2080 Ti. But I expect the price of the 2080 Ti will come down while we wait for the code to catch up.

Then we should arrange a little fund raising campaign for the GPUGRID project to be able to hire a CUDA programmer who'll update the application.

I'd like to see some financial support for those of us that provide client support services for all distributed computing projects. It used to be that CURE, FLDC & GRC were worth more and we could sell them to cover electricity. Today they need to be 5x to 10x higher to play the electric bill.

Jim1348
Send message
Joined: 28 Jul 12
Posts: 696
Credit: 1,371,999,968
RAC: 54
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 51095 - Posted: 26 Dec 2018 | 15:27:34 UTC

It appears that Nvidia will offer some new cards (at least at the low end) without ray tracing. Maybe that will alleviate some of the compatibility problems? I would wait in any case to see though; that is good advice.
https://www.tomshardware.com/news/nvidia-geforce-gtx-2050-gtx-1150-specs,38291.html

Yes, it does take a long time to post here from the U.S., which is not unusual.

Aurum
Send message
Joined: 12 Jul 17
Posts: 118
Credit: 7,469,926,993
RAC: 3,552,085
Level
Tyr
Scientific publications
wat
Message 51096 - Posted: 26 Dec 2018 | 15:37:22 UTC - in response to Message 51018.
Last modified: 26 Dec 2018 | 15:37:51 UTC

...Turing has not improved much in regard to power efficiency compared to Pascal. Nvidia saved some GPU space by the shrink on one hand but used that saving for introduction of new Raytracing cores on the other hand.

The Folding@home fans say PPD/Watt went up from 4.8 for a 1080 Ti to 6.7 for a 2080 Ti. Not sure if comparing F@H to GPUgrid is apples and oranges.

I would love to see HCF1 protein folding and interaction simulations to help my little boy

I'd like to learn more about that. Is there a thread on this topic???

Aurum
Send message
Joined: 12 Jul 17
Posts: 118
Credit: 7,469,926,993
RAC: 3,552,085
Level
Tyr
Scientific publications
wat
Message 51097 - Posted: 26 Dec 2018 | 17:43:43 UTC - in response to Message 51086.

I started following some of the links they've posted. Is this the repository you asked about?
https://github.com/acellera/htmd
From http://www.compscience.org/

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 51098 - Posted: 26 Dec 2018 | 18:02:46 UTC - in response to Message 51097.
Last modified: 26 Dec 2018 | 18:06:55 UTC

Yes, the github repository is the one I was looking for. That is where our code for the apps reside. It is open-source software as required by the GPL license. The version we use is not the commercial pro version.

Now to find that snippet of code that produces the error.

[Edit] I'm not finding that text in the code base. It must be coming in from outside through a call to another module. Perhaps the conda environment which the app uses?

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 319,920
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51109 - Posted: 26 Dec 2018 | 23:07:51 UTC - in response to Message 51095.

It appears that Nvidia will offer some new cards (at least at the low end) without ray tracing. Maybe that will alleviate some of the compatibility problems?

I don't think so. As far as we the problem is not missing compatibility, but missing validation. And the new cards without RT cores are rumoared to also have Turing shader cores, so the new "CUDA compute capability" - which is tested for in the GPU-Grid code - also applies to them. Otherwise they'd simply be new Pascals with GDDR6, which wouldn't make sense for nVidia.

@Keith & Aurum: so there is a repository. Interesting! I don't have the knowledge nor time to dig into this, though :/

MrS
____________
Scanning for our furry friends since Jan 2002

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 51111 - Posted: 27 Dec 2018 | 1:39:28 UTC

I'm slowly getting to know how to probe github repositories. I've found the code I need to alter in the work_fetch.cpp module for the BOINC client for my older client.

I am waiting on the master code base changes for the fix I need to use the <gpu_exclude> and <max_concurrent> statements at the same time. I use the <gpu_exclude> for GPUGrid and Einstein to keep the RTX2080 card off those projects for now.

Richard and David are working on that bug that I started on the BOINC github repository.

It is going to be up to me to figure out how to compile the client for Linux.

Richard says the new client will be automatically generated for Windows by the appveyor application once they submit the bug fix and the request is pulled into master. That is pretty cool. One of the benefits of running Windows I guess.

JoergF
Avatar
Send message
Joined: 20 Apr 15
Posts: 283
Credit: 1,101,827,807
RAC: 111,337
Level
Met
Scientific publications
watwatwat
Message 51112 - Posted: 27 Dec 2018 | 9:02:45 UTC
Last modified: 27 Dec 2018 | 9:05:24 UTC

I would love to see HCF1 protein folding and interaction simulations to help my little boy


I'd like to learn more about that. Is there a thread on this topic???


Thanks for asking... no, not yet. My little son has a single point missense mutation in the basic region of HCF1 and nobody can tell you what this HUGE protein is supposed to do there. From the papers it seems that HCF1 is some kind of control protein in the cell cycle and for gene expression in combination with many others like SIN3A ... but most of it is not yet understood or investigated. For this kind of science it seems that we still are decades away in terms of computing power :(
____________
I would love to see HCF1 protein folding and interaction simulations to help my little boy... someday.

JoergF
Avatar
Send message
Joined: 20 Apr 15
Posts: 283
Credit: 1,101,827,807
RAC: 111,337
Level
Met
Scientific publications
watwatwat
Message 51113 - Posted: 27 Dec 2018 | 9:16:29 UTC

One more idea to solve the problems of having low funds at the GPUGRID institute for additional researchers and (therefore) lack of tasks (which has been improved recently).

When there are no jobs available, we could install a "bridge" to redirect the GPUGRID crunching power to CureCoin mining in support of the institute. The earnings of this mining shall be spent for additional headcounts, more researchers or CUDA programmers. As long as science prevails, it should be a good cause.
____________
I would love to see HCF1 protein folding and interaction simulations to help my little boy... someday.

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51114 - Posted: 27 Dec 2018 | 10:26:02 UTC - in response to Message 51097.

I started following some of the links they've posted. Is this the repository you asked about?
https://github.com/acellera/htmd
From http://www.compscience.org/

The present GPUGrid (GPU) app is acemd2. We're waiting for the acemd3 to be made available for GPUGrid. I think that acemd3 supports the latest NVidia cards, as it is based on OpenMM. I think that HTMD is CPU only.
https://software.acellera.com/docs/latest/index.html

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51115 - Posted: 27 Dec 2018 | 11:00:33 UTC - in response to Message 51113.
Last modified: 27 Dec 2018 | 11:14:45 UTC

JoergF wrote:
When there are no jobs available, we could install a "bridge" to redirect the GPUGRID crunching power to CureCoin mining in support of the institute. The earnings of this mining shall be spent for additional headcounts, more researchers or CUDA programmers. As long as science prevails, it should be a good cause.

Aurum wrote:
I'd like to see some financial support for those of us that provide client support services for all distributed computing projects. It used to be that CURE, FLDC & GRC were worth more and we could sell them to cover electricity. Today they need to be 5x to 10x higher to play the electric bill.

I didn't check it, but I think that at the moment it's better (more green anyway) to send the donation directly in "real world" currencies than in "virtual" coins, as the electricity used for generating these costs more. But I'm aware that this old fashioned way of donation is less appealing for crunchers like us. There was an attempt by GPUGrid to do that (also to involve AMD card owners in this project) called donate@home. It didn't work back then. Since then I came to the conclusion (by understanding the way the late BitcoinUtopia project worked; it was similar to donate@home btw) that it's immoral to give (actually sell) BOINC credits for virtual (or real world) coins. GRidCoin and CureCoin does the opposite, so it's ok, but I don't think that a project's funding can depend on exchange rate booms, as it's like lottery. That's why I stick to:
Retvari Zoltan wrote:
Then we should arrange a little fund raising campaign for the GPUGRID project to be able to hire a CUDA programmer who'll update the application.

JoergF
Avatar
Send message
Joined: 20 Apr 15
Posts: 283
Credit: 1,101,827,807
RAC: 111,337
Level
Met
Scientific publications
watwatwat
Message 51116 - Posted: 27 Dec 2018 | 12:57:58 UTC
Last modified: 27 Dec 2018 | 13:05:02 UTC

Zoltan, I really appreciate your proposal. However I think we must take a couple of organizational things into consideration as well. Some time ago, I have made a suggestion relative to students getting familiar with this kind of science. The idea was to provide GPUGRID computing power for laboratory tutorials, which eventually lead to PhD, diploma or project works to take a little pressure off the GPUGRID researchers.

But as Stefan pointed out, it apparently is not so easy to catch interest in this matter, train the students and get valuable results in a short time. See this post/topic:

http://www.gpugrid.net/forum_thread.php?id=4443&nowrap=true#46311

So I understand the institute might rather be interested in hiring additional researchers at fixed salary which requires a permanent flow of money. Unfortunately a single fund raising campaign would not do that. So I was trying to think outside the box and find some permanent income for them, although there will be fluctuations, as you wrote.

PS: why did the BitcoinUtopia project fail? The underlying idea doesn't sound so bad.
____________
I would love to see HCF1 protein folding and interaction simulations to help my little boy... someday.

Aurum
Send message
Joined: 12 Jul 17
Posts: 118
Credit: 7,469,926,993
RAC: 3,552,085
Level
Tyr
Scientific publications
wat
Message 51121 - Posted: 27 Dec 2018 | 17:59:34 UTC
Last modified: 27 Dec 2018 | 18:00:57 UTC

I was thinking that these guys really seem to know what they're doing. There must be labs that need a steady diet of molecular simulations but don't have the resources to set up the code and servers. Could GDF provide this service to other researchers for a fee??? It would increase the supply of projects & WUs to "Feed the Beast" as the OpenZika folks say.

I thought about submitting a proposal to The Gates Foundation but they only accept them from non-profit orgs like the @CureCoin_Team. Is Gridcoin Network a non-profit org? Sue Hellmann MD PhD (my alma mater) used to run a biotech so she knows the value of molecular simulations.

I dream of having a solar-powered DC Client Center.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 319,920
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51122 - Posted: 27 Dec 2018 | 22:33:39 UTC - in response to Message 51121.

Universities are non-profit by definition. They may offer their services to customers, but will reinvest the money earned directly to aid their non-profit goals.

MrS
____________
Scanning for our furry friends since Jan 2002

Keith Myers
Send message
Joined: 13 Dec 17
Posts: 296
Credit: 239,043,213
RAC: 21,142
Level
Leu
Scientific publications
wat
Message 51123 - Posted: 27 Dec 2018 | 22:42:33 UTC - in response to Message 51114.

I started following some of the links they've posted. Is this the repository you asked about?
https://github.com/acellera/htmd
From http://www.compscience.org/

The present GPUGrid (GPU) app is acemd2. We're waiting for the acemd3 to be made available for GPUGrid. I think that acemd3 supports the latest NVidia cards, as it is based on OpenMM. I think that HTMD is CPU only.
https://software.acellera.com/docs/latest/index.html

Thanks for that link Zoltan. I think you are correct in that we will need to wait for the acemd3 application to be released.

Also I think some of the installation instructions I read in those docs reinforce my belief that the conda environment is probably the one that is causing the error outputs in stderr.txt.

Aurum
Send message
Joined: 12 Jul 17
Posts: 118
Credit: 7,469,926,993
RAC: 3,552,085
Level
Tyr
Scientific publications
wat
Message 51125 - Posted: 28 Dec 2018 | 2:13:15 UTC - in response to Message 51122.
Last modified: 28 Dec 2018 | 2:13:58 UTC

Universities are non-profit by definition. They may offer their services to customers, but will reinvest the money earned directly to aid their non-profit goals. MrS


The foundation awards the majority of its grants to U.S. 501(c)(3) organizations and other tax-exempt organizations identified by our staff. (Tax status definitions) (Glossary of terms)

https://www.gatesfoundation.org/How-We-Work/General-Information/Grant-Opportunities

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 319,920
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51133 - Posted: 28 Dec 2018 | 20:55:27 UTC - in response to Message 51125.

The only passage about university I found is

Foreign Government
Your organization is a political subdivision, agency or instrumentality of a foreign national, state or local government (e.g., foreign public university, foreign ministry of health or education).

... does that mean thumbs up or down?

MrS
____________
Scanning for our furry friends since Jan 2002

Profile bcavnaugh
Send message
Joined: 8 Nov 13
Posts: 54
Credit: 830,150,450
RAC: 3,149,195
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwat
Message 51144 - Posted: 30 Dec 2018 | 2:09:34 UTC

Thanks Penguin
I will say the RTX run PG and Collatz Really Well.
We still have 3 or 4 BOINC Projects that are not yet supporting RTX but in Time I hope they do.

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1919
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 51146 - Posted: 30 Dec 2018 | 8:42:38 UTC - in response to Message 51144.

We will support RTX. Next year I will try to find a better solution for gpugrid.
At the moment some people in the lab are using it but nobody is really responsible for it or developing it. Ideally we would need somebody with very good programming skills in C/C++ and knowledge of CUDA, IT admin and high performance computing.

gdf

Profile Retvari Zoltan
Avatar
Send message
Joined: 20 Jan 09
Posts: 2054
Credit: 15,015,899,469
RAC: 7,808,568
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51206 - Posted: 7 Jan 2019 | 12:12:51 UTC - in response to Message 51199.

Does this project use double precision on a GPU...?
No, it doesn't.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2688
Credit: 1,172,901,099
RAC: 319,920
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 51729 - Posted: 26 Apr 2019 | 21:14:52 UTC
Last modified: 8 Jun 2019 | 20:45:05 UTC

I'm pinning and locking this thread, so people can more easily find GDFs answer on Turing support.

Update: see there for new developments:
Update 2: it's progressing! [link]

MrS
____________
Scanning for our furry friends since Jan 2002

Message boards : Graphics cards (GPUs) : Turing support status at end of thread (former RTX 2080 TI thread)