I have rig with 4 * 1080ti
And Intel(R) Core(TM) i3-8100 CPU @ 3.60GHz [Family 6 Model 158 Stepping 11] (4 core)
But looks like calculation very slow ~ 8-10 hour for long stats:
http://www.gpugrid.net//results.php?hostid=471190

But i see people with 1080ti too and spend of the same tasks ~2-3h.
http://www.gpugrid.net/results.php?hostid=458167

What i am doing wrong.

I have linux ubuntu 17.10.


Some ideas:

- No multiple GPUs in the same system to achieve PCIe3.0x16.

PCIe3.0x16 required ?

Now card connected via:

https://www.aliexpress.com/item/New-3-in-1-Molex-4Pin-SATA-6PIN-PCIE-PCI-E-PCI-Express-16x-Slot-Riser/32838716482.html?spm=a2g0s.9042311.0.0.W0tvgJ

Hello, unlike normal cryptocurrency mining, some scientific applications requires PCIe bandwidth. From my experience something like the application for Einstein@home or SETI@home require the least PCIe bandwidth. A 1x link will do just fine. If you would like me to go further into optimizing please PM me.

If you wish to contribute to this exact project, this application requires large PCIe bandwidth, which means 16x riser cables and the infrastructure to support it (aka a CPU that has at least 8x PCIe x3 slots for each high speed GPU.)

One of my machines I have an i3-7100 with only one 1080ti to get the maximum speed out of the card. I can achieve max speed because there is an entire 16x PCIe slot dedicated to this one card which is the entirely of the i3-7100's PCIe bandwidth. The double precision calculations are done on the CPU so PCIe bandwidth has to be very high.

If you wish to contribute to this project I suggest spreading out your GPUs among other systems to maximize PCIe bandwidth for each card.

As for operating system you have gone with the right choice, Linux is amazing for most applications, but especially for science. Many scientific applications run significantly faster on Linux and some applications are ONLY written for Linux. Some projects have to use Virtual Machines to give work to windows users, at huge computational loss. For example the Large Hadron Collider project's application is only written for Linux without the use of a VM.

If you wish to support this project, please keep in mind that the work is not guaranteed and often the network and compute their work faster than they can give new work, leading to shortages fairly often. I would have a backup project that BOINC automatically switches to if there is no GPUGrid work.

Since you are on Linux you can however help out the CPU workload for GPUGrid which requires Linux and they are desperate to speed up the computational time.

PCIe3.0x16 required ?

Now card connected via:
https://www.aliexpress.com/item/New-3-in-1-Molex-4Pin-SATA-6PIN-PCIE-PCI-E-PCI-Express-16x-Slot-Riser/32838716482.html?spm=a2g0s.9042311.0.0.W0tvgJ

try buying an older generation xeon. I bet you can get 1650v3 fairly cheaply these days. These will givve you all the PCI lanes you need in a single rig.

http://www.gpugrid.net/forum_thread.php?id=4620#47847

I just say this thread so single thread appears to be an issue too. MAkes sense why my 980TI on my 4.2GHz 1650v3 wont fun at max speed.

You might want to buy a binned or at least a high clocked baby lake CPU.

Unfortunately high IPC, Clockspeed and PCIe bandwidth are required for fast GPUs like 980ti, 1080 and 1080ti for maximum speed. SWAN_SYNC is also required for maximum performance.

I agree to the below comments of PappaLitto and Zoltan. Which mainboard do you have, is it already a Z370 ... or still Union Point?

The first (Coffee Lake) and a few Union Point chipsets do already support 24 lanes of PCI_3.0 speed, that said, not all of them are available for GPU. Most mainboards on market support only 3 Cards at [x8/x8/x4] at the same time ... only the ASUS ROG Maximus X Apex seems to support four at [x8/x8/x4/x4] which is extraordinary.

No idea whether or not a 1080ti is bottlenecked at even PCIe_3.0x4, it may be, but I have not tried yet. Mine runs on a single PCIe 3.0x16 (both mechanical and electrical), leaving the other slots empty. But 2.0x1 definitely is inappropriate as Zoltan already wrote.

Either way, I assume that your mainboard can only host three 1080ti (run concurrently at x8/x8/x4) and also the i3-8100 seems to be a little slow for supporting these powerful cards. But we certainly can check that once you have revealed us your mainboard type/model.
____________
I would love to see HCF1 protein folding and interaction simulations to help my little boy... someday.

My motherboard:
https://www.asus.com/Motherboards/PRIME-Z370-P/

But unfortunately i can not setup even 2 1080ti into one motherboard because it will be very hot temperature in card.

i can not setup even 2 1080ti into one motherboard because it will be very hot temperature in card