Advanced search

Message boards : Graphics cards (GPUs) : PCIE Riser?

Author Message
Profile Francois Normandin
Send message
Joined: 8 Mar 11
Posts: 71
Credit: 638,327,551
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwat
Message 39701 - Posted: 25 Jan 2015 | 18:57:55 UTC
Last modified: 25 Jan 2015 | 19:08:45 UTC

Did using PCIE Riser change something on the performance? Using x1?

Profile MJH
Project administrator
Project developer
Project scientist
Send message
Joined: 12 Nov 07
Posts: 696
Credit: 27,266,655
RAC: 0
Level
Val
Scientific publications
watwat
Message 39706 - Posted: 25 Jan 2015 | 20:55:55 UTC - in response to Message 39701.

Using one is likely to increase your error rate somewhat. They're not very reliable.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2674
Credit: 910,467,424
RAC: 553,058
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 39708 - Posted: 25 Jan 2015 | 22:02:03 UTC

Depending on the performance of your GPU and PCIe version used 1x is really slow. 4x PCIe 3 is still approximately enough for the current GPU-Grid app, though.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,924,393,960
RAC: 199,068
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 39742 - Posted: 27 Jan 2015 | 9:45:19 UTC - in response to Message 39706.
Last modified: 27 Jan 2015 | 9:50:20 UTC

Using one is likely to increase your error rate somewhat. They're not very reliable.

From past experience I would agree, but you can keep the slightly higher error rate down by using a good PSU, not overclocking, and running cool (which is easier anyway).
Well made (copper wire/connector) risers are ok, but there are some low-grade (cheap alloy) versions. If it looks as cheap as tin, it is.
As GPU's take power from the PCIE slot it's a good idea to get a riser with a power cable (often molex). This covers the PCIE power loss from the cable length and additional connection.
Overclocking on a riser would be pushing it, but I expect the Maxwell's are a lot more power resilient and reliable on a riser than the Keplers and previous generations.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile skgiven
Volunteer moderator
Project tester
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,924,393,960
RAC: 199,068
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 40677 - Posted: 29 Mar 2015 | 13:41:21 UTC - in response to Message 39742.
Last modified: 7 Jun 2015 | 8:53:22 UTC

Experiment to assess using PCIEx1.

Ran an e1s5_2-GERARD_FXCXCL12_LIG_11543842-0-1-RND7987_0 GPUGrid task on a GTX670 mounted via a molex powered PCEIx1 to PCIEx16 riser.



These GERARD_FXCXCL12 tasks are very long and it took 64,881 sec (18h) to complete. I got full credit, because my cache just happened to be low enough, but I had less than 1h to spare:
27 Mar 2015 | 16:10:19 UTC 28 Mar 2015 | 15:21:59 UTC Completed and validated 64,881.79 64,286.02 255,000.00

For comparison my GTX970 took 13.75h:
27 Mar 2015 | 12:13:29 UTC 28 Mar 2015 | 6:31:38 UTC Completed and validated 49,556.33 49,170.73 255,000.00

Setup:
W7x64 (WDDM overhead), also running Einstein on my integrated gpu (significant overhead there too), and 5 CPU tasks (impacts somewhat on performance, a few percent at least). The GTX670 (170W TDP) was downclocked by 65MHz to stay closer to the 2*6-pin (150W) power margins, though the molex power connector could in theory supply 12V*5A=60W. It ran at 95% power according to MSI Afterburner.

For comparison reasons I was not hoping for one of these new extra-long tasks, but others report they run slightly hotter and they do run much longer, so it was more likely to fail than the usual long tasks.

Although I'm also not directly able to assess the relative performance loss (PCIEx1 vs PCIEx16) due to running multiple CPU and iGPU tasks, and limited task history, it did test proof of concept in a stressful environment. Performance was ~18% slower than optimal (going by approx. relative performances of the 970 and 670) but it should be possible to get within 15% of normal performance (for that card).
Note that if the GPU was actually in the system it would be a second GPU anyway and would thus be operating at PCIE 3.0 X8 (not PCIE 3.0 X16). A third GPU would also be operating at PCIE 2 X8 or X4. So any small difference between X8 or X4 and X16 should be considered.

While using PCIeX1 worked reasonably well (no reported task recovery), the exact loss due to PCIEx1 is not known. From experience, it's likely that the loss scales with GPU performance and varies by task type. So the loss on a GTX750Ti would for example be around about half the loss on a GTX970 and for some tasks it would be larger than others (the smaller the atom count in the model the greater the loss would be, I think).
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Post to thread

Message boards : Graphics cards (GPUs) : PCIE Riser?