GTX 1080 time per task difference
log in

Advanced search

Message boards : Windows : GTX 1080 time per task difference

Author Message
Profile Grigory Kostykov
Send message
Joined: 7 Oct 12
Posts: 19
Credit: 3,429,137,042
RAC: 0
Message 23450 - Posted: 28 Nov 2016, 0:59:23 UTC

I have 3 PC.
1-I7+1080
2-I3+1080
3-G4500+2x1080
(all 1080 is identical)
On first 2 pc collatz task takes 1m23s-1m27s to complete.
On 3rd pc it takes 1m37s on both cards. (Mb is Asus z170-a, 8gb 3000mhz DDR4.)
When i look on GPU load monitor on 1 pc and 2 pc, gpu have 97-98% utilization. On 3rd pc both GPU have only 85-90% utilization. PC is clean, no any other task running.
Something like this was on 2 pc after i flash new bios, but after reflash old one, time of task returned to normal.I did try all versions of bios on 3 pc didn't help.
I think there can be reasons:
1. CPU - change to more powerful (but CPU load is 1-2%)
2. MB - ? (just like it, and only mb from other vendor can fix it)
3. Config - ? (some special tuning)

Any suggestion?

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 2
Message 23454 - Posted: 28 Nov 2016, 4:16:01 UTC

When running multiple PCI Express cards, each additional card may cause the transfer speed to be reduced. For example, a 16x PCI Express slot may drop to 8x if a second PCI Express slot is also populated. In other words, 16/2 = 8. Normally, that doesn't affect most programs since they seldom reach or can even come close to reaching the max throughput. For the Collatz sieve workunits, a 32-bit sieve requires a huge amount of memory and the bottleneck may potentially be an issue. You might want to see if the machine with dual 1080 GPUs would be faster with a slightly lower sieve size.


As I shoot myself in the foot..... The largest sieve size a GPU can process doesn't always result in the max credits that the GPU can achieve running Collatz. While it would be nice if that were the case as that results in the max science being done, there are times where a smaller sieve size (26-30) results in more credits per hour for the same GPU. The smaller sieve size means that more numbers are checked resulting in more total steps. Since credit is determined by total steps, it is possible to get more total steps in less time with slightly lower sieve sizes. It generally means the GPU will also run hotter. If heat gets too high, it will slow itself down automatically. My laptop can't run over 90% without throttling the GPU speed. When it does, the load remains at 99% but the actual GPU speed is reduced. Hope that helps.

Profile Grigory Kostykov
Send message
Joined: 7 Oct 12
Posts: 19
Credit: 3,429,137,042
RAC: 0
Message 23457 - Posted: 28 Nov 2016, 15:02:09 UTC - in response to Message 23454.

Thanks. And yes i know about all this PCI x16 architecture, how it splits and counts etc.

I remove one card and run task - results is same. So PCIx lines not affected bcz it was directy connected to cpu by his built-in PCIx lines. And i can hardly belive that collatz will utilize 4gb\sec(x16) 2gb\sec(x8) 1gb\sec(x4) even 500mb\sec for 2x bcs they a bi-directed so it means 8-4-2-1 Gb\sec.
Task takes around 75-80 mb of ram so not affected by ram size.
Also task takes around 140mb of videocard memory so also not affected by it.

But i get noticed when slider of % task moves it got some frizes around 5-7 times. Fast ones like 1 sec. But on other PC's it not happening. Like something lagging.

Still got not fully loaded GPU's. I see one exit from this: to buy fast CPU like I5-K 4ghz (maybe it will affect on GPU's load), and addind more RAM (to use both dual channels, maybe it will afeect on chipset architecture CPU-PCIe-Chipset data exchange). Or maybe it just motherboard ! But it one of the best from Asus.

And what you mean by "would be faster with a slightly lower sieve size" ?

Profile Slicker
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist
Avatar
Send message
Joined: 11 Jun 09
Posts: 2525
Credit: 740,580,099
RAC: 2
Message 23458 - Posted: 28 Nov 2016, 16:07:09 UTC
Last modified: 28 Nov 2016, 16:07:22 UTC

Have you read the thread that explains how to optimize the sieve app?

https://boinc.thesonntags.com/collatz/forum_thread.php?id=1273#20542

Profile Grigory Kostykov
Send message
Joined: 7 Oct 12
Posts: 19
Credit: 3,429,137,042
RAC: 0
Message 23461 - Posted: 28 Nov 2016, 17:22:48 UTC - in response to Message 23458.

Thanks !
Reading and experimenting now.
Got 45 sec per WU !
If my calculations is right, i will got +2.2 million of points on each card !


Post to thread

Message boards : Windows : GTX 1080 time per task difference


Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.