Advanced search

Message boards : Number crunching : PCIe x1 risers

Author Message
PappaLitto
Send message
Joined: 21 Mar 16
Posts: 511
Credit: 4,672,242,755
RAC: 0
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 47825 - Posted: 29 Aug 2017 | 0:58:06 UTC
Last modified: 29 Aug 2017 | 0:59:01 UTC

So obviously, this project with its current application is a horrible candidate for GPUs connected with a pcie x1 riser connection. I tried it on a 970 just in case anyone was curious what would happen. The BUS interface load was 80% and the GPU load was 69% but even with 69% it was much slower than a normal 69% GPU load. Results are here: https://www.gpugrid.net/result.php?resultid=16486751

My point is, does anyone know a project where I could run my 6 GPU system with pcie x1 risers full bore on? Maybe einstein@home?

[CSF] Aleksey Belkov
Send message
Joined: 26 Dec 13
Posts: 86
Credit: 1,283,607,256
RAC: 347,000
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 47826 - Posted: 29 Aug 2017 | 8:59:56 UTC - in response to Message 47825.

Might be worth a try at Folding@HOME?
It's not BOINC project, but biological research much better usage of electricity.

kain
Send message
Joined: 3 Sep 14
Posts: 152
Credit: 888,303,032
RAC: 1,175,837
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 47828 - Posted: 29 Aug 2017 | 10:28:34 UTC

I have tested folding and Einstein. Folding 70-75% of gpuload, Einstein - one GPU 70%, one 10%, one freezes for minutes... Risers 1x are generally bad for us.

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 511
Credit: 4,672,242,755
RAC: 0
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 47829 - Posted: 29 Aug 2017 | 10:49:41 UTC

When I tested Einstein, I got pulsing between 70 and 95% load with very little PCIe BUS interface load. This project reserves an entire thread on the CPU per GPU so make sure you have enough CPU power to accommodate the number of GPUs you have.

Erich56
Send message
Joined: 1 Jan 15
Posts: 1142
Credit: 11,053,436,328
RAC: 20,808,497
Level
Trp
Scientific publications
watwatwatwatwatwatwatwatwat
Message 47830 - Posted: 29 Aug 2017 | 12:39:34 UTC - in response to Message 47828.

I have tested folding and Einstein. Folding 70-75% of gpuload, Einstein - one GPU 70%, one 10%, one freezes for minutes...

on my main system, Einstein freezes right after start of a task, even without risers :-(

PappaLitto
Send message
Joined: 21 Mar 16
Posts: 511
Credit: 4,672,242,755
RAC: 0
Level
Arg
Scientific publications
watwatwatwatwatwatwatwat
Message 47832 - Posted: 29 Aug 2017 | 18:34:31 UTC

Did you try resetting the project?

WPrion
Send message
Joined: 30 Apr 13
Posts: 96
Credit: 3,289,071,611
RAC: 14,029,208
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 47852 - Posted: 6 Sep 2017 | 12:36:10 UTC
Last modified: 6 Sep 2017 | 12:44:27 UTC

I'm waiting for the 8 pcie-x8 slot motherboard for Threadripper to make use of it's 64 pci-e lanes. Then there's the EPYC with 128 lanes... ;-)

Sad, though, there are several readily available boards with up to 19 pcie-x1 slots for bitcoin etc., but for science, no so much.

edited to add:
I see 4x risers on the market and many boards have 7 slots. Has anybody tried those?

WPrion
Send message
Joined: 30 Apr 13
Posts: 96
Credit: 3,289,071,611
RAC: 14,029,208
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 47867 - Posted: 13 Sep 2017 | 12:58:11 UTC - in response to Message 47852.
Last modified: 13 Sep 2017 | 13:05:51 UTC

speaking of...but still mostly x1 slots:

http://www.pcgamer.com/heres-an-insanely-long-motherboard-with-8-pcie-x16-slots/

W

wiyosaya
Send message
Joined: 22 Nov 09
Posts: 114
Credit: 589,114,683
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 47868 - Posted: 13 Sep 2017 | 15:55:19 UTC - in response to Message 47867.
Last modified: 13 Sep 2017 | 15:57:55 UTC

speaking of...but still mostly x1 slots:

http://www.pcgamer.com/heres-an-insanely-long-motherboard-with-8-pcie-x16-slots/

W

Marketing! :(

As for Threadripper, I have seen MBs where all x16 slots are not electrically x16 slots. It looks like the MSI has 4, electrically x16 slots, however, it would be worth an e-mail to them if one were going to buy this board to ensure that they are all x16 slots. https://www.msi.com/Motherboard/X399-GAMING-PRO-CARBON-AC

There is an EPYC proc that is about the same price as the 1950X, however, the MBs out there as of now do not look like they would be good choices in which to run 8 gpus - IMO.
____________

WPrion
Send message
Joined: 30 Apr 13
Posts: 96
Credit: 3,289,071,611
RAC: 14,029,208
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 47885 - Posted: 16 Sep 2017 | 14:42:04 UTC - in response to Message 47868.

Even when slots are capable of x16 data you have to read the fine print on configuration. Often, adding an M.2 pcie ssd steals x4 lanes from somewhere. And since there are no x12 slots one of the x16 slot will drop to x8 or an x8 slot will be disabled.

Win

oemuser
Send message
Joined: 18 Sep 16
Posts: 10
Credit: 1,291,979
RAC: 0
Level
Ala
Scientific publications
wat
Message 47893 - Posted: 18 Sep 2017 | 11:28:07 UTC

folding@home with x1 risers Windows bottlenecks pcie bandwidth but on Linux it is working much better. Maybe the same is true for gpugrid?

Post to thread

Message boards : Number crunching : PCIe x1 risers

//