| Author |
Message |
|---|
|
|
|
Would love to see some speeds from the new 1080 |
|
|
Monkeydee Volunteer tester
Send message
Joined: 8 Dec 13
Posts: 540
ID: 284516
Credit: 1,531,877,060
RAC: 820,374
                            
|
|
For $950 from Newegg in Canada I think I will be taking a pass at one.
However, major price drops on the 980 and 980ti make them kinda tempting.
Very curious to see some primegrid based performance numbers though.
____________
My Primes
Badge Score: 4*2 + 6*2 + 7*3 + 8*10 + 11*3 + 12*1 = 166
 |
|
|
|
|
|
Some preliminary synthetic compute benchmarks for the GTX 1080 using early release GeForce driver. |
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
                                                    
|
|
Just managed to order one from Newegg.
Hopefully, will have some numbers next week.
____________
 |
|
|
|
|
|
Think I'll pick one up in 2020, right after the 2080s come out. Should cost about $150 USD by then. My entire grid is not worth $1000. Joking aside, also interested in seeing the speeds these monsters produce. |
|
|
|
|
|
http://www.primegrid.com/show_host_detail.php?hostid=465745
PPS (Sieve): 270sec
Trying GFN=22 with geneferocl4.
It's completed in 153000 sec in the plan.
About 11% faster than GTX TITAN X (173000sec). |
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
http://www.primegrid.com/show_host_detail.php?hostid=465745
PPS (Sieve): 270sec
Trying GFN=22 with geneferocl4.
It's completed in 153000 sec in the plan.
About 11% faster than GTX TITAN X (173000sec).
Can't wait for mine to get here.
If it's that fast, I'll probably buy more -- faster + lower power requirements is very attractive as summer comes on.
____________
|
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
|
1080 up and running on this host.
Check that host for BOINC subproject results. I will report results on manual sieving.
Looks great on PPS Sieve.
____________
|
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
1080 up and running on this host.
Check that host for BOINC subproject results. I will report results on manual sieving.
Looks great on PPS Sieve.
Just switched to manual sieving -- running a GFN N=2097152 task at 124P/day.
Makes the 200P reservation limit seem a little small.
____________
|
|
|
|
|
|
GFN=22 test finished.
http://www.primegrid.com/result.php?resultid=716883642
149,041.65sec
|
|
|
 RafaelVolunteer tester
Send message
Joined: 22 Oct 14
Posts: 911
ID: 370496
Credit: 551,549,092
RAC: 448,706
                   
|
|
41 hours.... 1 and 2/3 days......
Wow.
Just wow.
Ofc manual sieving is still faster, but still.... wow. This is fast. Is it OCed or at stock? |
|
|
|
|
|
Stock FE model, but frequency rose automatically and became 1848MHz.
The temperature was stable at 68°C with fan level 85%. (Room 26°C) |
|
|
GDBSend message
Joined: 15 Nov 11
Posts: 298
ID: 119185
Credit: 4,066,754,817
RAC: 1,926,070
            
|
1080 up and running on this host.
Check that host for BOINC subproject results. I will report results on manual sieving.
Looks great on PPS Sieve.
Just switched to manual sieving -- running a GFN N=2097152 task at 124P/day.
Makes the 200P reservation limit seem a little small.
What B are you using? How many GPU tasks are you running? How busy is the GPU?
124P/day for n=21 seems a little slow. I get 55P/day with a 660TI for n=22 running 2 tasks @ B=11. GPU is 95-99% busy. I would expect 1080 to be at least 3 times faster than 660TI. |
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
1080 up and running on this host.
Check that host for BOINC subproject results. I will report results on manual sieving.
Looks great on PPS Sieve.
Just switched to manual sieving -- running a GFN N=2097152 task at 124P/day.
Makes the 200P reservation limit seem a little small.
What B are you using? How many GPU tasks are you running? How busy is the GPU?
124P/day for n=21 seems a little slow. I get 55P/day with a 660TI for n=22 running 2 tasks @ B=11. GPU is 95-99% busy. I would expect 1080 to be at least 3 times faster than 660TI.
B=11, 1 task.
B=13 with 1 task is slightly faster, and there is no noticeable screen lag.
Not really seriously testing manual sieving at this point -- just finishing up the task that was running on that system. Will do more serious testing once I have gone through the BOINC subprojects.
____________
|
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
|
My host with the 1080 Now has results for BOINC subprojects up to GFN-20. 21 and 22 will have to wait until I finish testing manual sieving.
Running 2 GFN 2097152 manual sieving tasks with blocking factor 13, I'm seeing 77P/day each, or 154P/day total. EVGA Precison X says the GPU temp is 78C. Card fan is at max, but case fans are not.GPU clock is running at 1733 MHz.
____________
|
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
My host with the 1080 Now has results for BOINC subprojects up to GFN-20. 21 and 22 will have to wait until I finish testing manual sieving.
Running 2 GFN 2097152 manual sieving tasks with blocking factor 13, I'm seeing 77P/day each, or 154P/day total. EVGA Precison X says the GPU temp is 78C. Card fan is at max, but case fans are not.GPU clock is running at 1733 MHz.
Addendum: running these tasks, there is a slight screen lag, but the desktop is still quite usable (with a 4K and a 2560x1440 monitor attached). System has 8 physical cores, HT off, and is running 8 LLR tasks in addition to the manual sieving.
Ooops-- it's also running a GFN-17 Low BOINC task -- will update when that completes.
____________
|
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
My host with the 1080 Now has results for BOINC subprojects up to GFN-20. 21 and 22 will have to wait until I finish testing manual sieving.
Running 2 GFN 2097152 manual sieving tasks with blocking factor 13, I'm seeing 77P/day each, or 154P/day total. EVGA Precison X says the GPU temp is 78C. Card fan is at max, but case fans are not.GPU clock is running at 1733 MHz.
Addendum: running these tasks, there is a slight screen lag, but the desktop is still quite usable (with a 4K and a 2560x1440 monitor attached). System has 8 physical cores, HT off, and is running 8 LLR tasks in addition to the manual sieving.
Ooops-- it's also running a GFN-17 Low BOINC task -- will update when that completes.
GFN-17 task complete -- manual sieves now running at 78P/day -- 156P/day toatl.
____________
|
|
|
Yves GallotVolunteer developer
Project scientist
Send message
Joined: 19 Aug 12
Posts: 820
ID: 164101
Credit: 305,989,513
RAC: 2,326
|
My host with the 1080 Now has results for BOINC subprojects up to GFN-20. 21 and 22 will have to wait until I finish testing manual sieving.
Then we have:
28002178^32768+1 49 sec 810648 bits => 60 us/mul
12982154^65536+1 167 sec 1548617 bits => 108 us/mul
4438596^131072+1 320 sec 2894288 bits => 110 us/mul
1716640^262144+1 951 sec 5429305 bits => 175 us/mul
1107810^524288+1 3840 sec 10527325 bits => 365 us/mul
766104^1048576+1 14130 sec 20496704 bits => 689 us/mul
? => 1100/1200 us/mul ?
70862^4194304+1 149270 sec 67581665 bits => 2200 us/mul
If N < 262144, GPU load < 100%. There are not enough tasks for 2560 cores!
A 1080 is able to square a 10 million digit number in about 1 millisecond...!!!
|
|
|
|
|
|
Hi, one my 980Ti get on Pps sieve for 300-320 sec. average .. custom nv 1080 card could maybe get to 200+
on the new titan and 1080ti will be even better.)) |
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
My host with the 1080 Now has results for BOINC subprojects up to GFN-20. 21 and 22 will have to wait until I finish testing manual sieving.
Running 2 GFN 2097152 manual sieving tasks with blocking factor 13, I'm seeing 77P/day each, or 154P/day total. EVGA Precison X says the GPU temp is 78C. Card fan is at max, but case fans are not.GPU clock is running at 1733 MHz.
Addendum: running these tasks, there is a slight screen lag, but the desktop is still quite usable (with a 4K and a 2560x1440 monitor attached). System has 8 physical cores, HT off, and is running 8 LLR tasks in addition to the manual sieving.
Ooops-- it's also running a GFN-17 Low BOINC task -- will update when that completes.
GFN-17 task complete -- manual sieves now running at 78P/day -- 156P/day toatl.
Now running two GFN N=4194304 manual sieving tasks at 84P/day each -- 168P/day total.
Oh yeah -- blocking factor 13. no noticeable screen lag.
____________
|
|
|
RafaelVolunteer tester
Send message
Joined: 22 Oct 14
Posts: 911
ID: 370496
Credit: 551,549,092
RAC: 448,706
|
Now running two GFN N=4194304 manual sieving tasks at 84P/day each -- 168P/day total.
709k PPD...
Hm, yummy... |
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
Now running two GFN N=4194304 manual sieving tasks at 84P/day each -- 168P/day total.
709k PPD...
Hm, yummy...
Yeah, my previous best card (GTX780ti) could do 95P/day on GFN 2097152 manual sieving tasks (running one task -- didn't do much better on throughput with two tasks). 50% performance increase.
____________
|
|
|
tngSend message
Joined: 29 Aug 10
Posts: 486
ID: 66603
Credit: 47,437,744,099
RAC: 27,430,125
|
|
Datum: a 5T PPR12M manual sieve in 46 hours (12345T-12350T).
Running two doesn't appear to improve throughput.
There were some other tasks running a small part of the time. Now running one with nothing else on the GPU.
____________
|
|
|
|
|
GFN=22 test finished.
http://www.primegrid.com/result.php?resultid=716883642
149,041.65sec
I think I'm going to actually sell a kidney on the black market, in order to get a 1080. I mean we only need one kidney, so technically it's overkill to have two. |
|
|
RafaelVolunteer tester
Send message
Joined: 22 Oct 14
Posts: 911
ID: 370496
Credit: 551,549,092
RAC: 448,706
|
GFN=22 test finished.
http://www.primegrid.com/result.php?resultid=716883642
149,041.65sec
I think I'm going to actually sell a kidney on the black market, in order to get a 1080. I mean we only need one kidney, so technically it's overkill to have two.
Or, you could get a 1070 + RX 480 for the price of a 1080 and give us benchmarks for both ;P |
|
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 433,660,774
RAC: 817,079
            
|
GFN=22 test finished.
http://www.primegrid.com/result.php?resultid=716883642
149,041.65sec
I think I'm going to actually sell a kidney on the black market, in order to get a 1080. I mean we only need one kidney, so technically it's overkill to have two.
True. With two kidneys, you're beyond "the optimal sieving point".
(Sorry, I just couldn't resist.)
____________
My lucky number is 75898524288+1 |
|
|
MonkeydeeVolunteer tester
Send message
Joined: 8 Dec 13
Posts: 540
ID: 284516
Credit: 1,531,877,060
RAC: 820,374
|
Or, you could get a 1070 + RX 480 for the price of a 1080 and give us benchmarks for both ;P
A local retailer just got a bunch of the RX 480's in today. I was tempted to go get one. But at $400CAD I might wait a bit.
Also, having to muck about with the app_info.xml business kind of puts me off getting any AMD card since my primary use has shifted from gaming to PrimeGrid.
____________
My Primes
Badge Score: 4*2 + 6*2 + 7*3 + 8*10 + 11*3 + 12*1 = 166
|
|
|
|
|
|
For those of you running 1080's, how loud are they?
I have an ASUS GTX 970 Strix and I can't even hear it when it's running anything from PrimeGrid.
But my other PC with a GTX 690 sounds like a hair dryer so I want to upgrade it with a 1080 if it's quiet.
Thanks |
|
|
mackerelVolunteer tester
Send message
Joined: 2 Oct 08
Posts: 2645
ID: 29980
Credit: 568,565,361
RAC: 198
              
|
|
Could just wait for the 1080 Strix?
In other news, the 1060 has been announced with availability from later this month. In CUDA cores it is half a 1080, but with more ram bandwidth per core. Should eventually be less than half the price, but not half the power. |
|
|
MonkeydeeVolunteer tester
Send message
Joined: 8 Dec 13
Posts: 540
ID: 284516
Credit: 1,531,877,060
RAC: 820,374
|
In other news, the 1060 has been announced with availability from later this month. In CUDA cores it is half a 1080, but with more ram bandwidth per core. Should eventually be less than half the price, but not half the power.
http://www.guru3d.com/news-story/nvidia-announces-geforce-gtx-1060.html
Interesting that they opted not to allow SLI on this card. They must have figured that people would buy two of these instead of one 1080 given that two 1060s would theoretically have similar performance to a 1080 and cost less.
Now I'm torn on an RX480 vs a GTX1060. The 480 looks good from a performance perspective for the price and the 1060 looks good from a power consumption perspective, while hopefully maintaining more than adequate performance, for the price.
No matter which decision I make I'm going to regret getting one over the other. The real world performance numbers on the 1060 and local pricing may sway that argument in the coming weeks.
____________
My Primes
Badge Score: 4*2 + 6*2 + 7*3 + 8*10 + 11*3 + 12*1 = 166
|
|
|
RafaelVolunteer tester
Send message
Joined: 22 Oct 14
Posts: 911
ID: 370496
Credit: 551,549,092
RAC: 448,706
|
Now I'm torn on an RX480 vs a GTX1060. The 480 looks good from a performance perspective for the price and the 1060 looks good from a power consumption perspective, while hopefully maintaining more than adequate performance, for the price.
No matter which decision I make I'm going to regret getting one over the other. The real world performance numbers on the 1060 and local pricing may sway that argument in the coming weeks.
AMD has problems with primegrid and Boinc (until Michael can fix both the server and client ends). And NVIDIA offers CUDA to use with GFN Manual Sieve (until JimB spits out the OCL app). If I was to buy a card right now, I'd say the 1060 is the preferred choice. |
|
|
MonkeydeeVolunteer tester
Send message
Joined: 8 Dec 13
Posts: 540
ID: 284516
Credit: 1,531,877,060
RAC: 820,374
|
AMD has problems with primegrid and Boinc (until Michael can fix both the server and client ends). And NVIDIA offers CUDA to use with GFN Manual Sieve (until JimB spits out the OCL app). If I was to buy a card right now, I'd say the 1060 is the preferred choice.
I don't bother with manual sieving. As for the rest of it, I could learn how to muck about with the app_info file to make it go. Others have been able to get their R9 3XX cards going, so I can't see an RX 4XX card being much different. But yes, the compatibility issues with here are definitely a consideration to be made.
____________
My Primes
Badge Score: 4*2 + 6*2 + 7*3 + 8*10 + 11*3 + 12*1 = 166
|
|
|
mackerelVolunteer tester
Send message
Joined: 2 Oct 08
Posts: 2645
ID: 29980
Credit: 568,565,361
RAC: 198
|
|
We could use more results in the AP27 test thread, but based on GPU results so far, nvidia cards seem to be pretty fast. (960 faster than fury nano).
There are plenty of arguments either way, some more primegrid, some more gaming. In another forum, someone declared the 1060 wasn't just slightly more expensive, but massively so because they wanted to go freesync/g-sync, and monitors with the latter costing much more than the former. |
|
|
|
|
We could use more results in the AP27 test thread, but based on GPU results so far, nvidia cards seem to be pretty fast. (960 faster than fury nano).
There are plenty of arguments either way, some more primegrid, some more gaming. In another forum, someone declared the 1060 wasn't just slightly more expensive, but massively so because they wanted to go freesync/g-sync, and monitors with the latter costing much more than the former.
That doesn't suprise me, the maxwell cards are pretty good at some single precision work. I haven't tested on my 980 yet, by my 290x was turning in 51-53 second times, and my 680s were turning in low-mid 90s. |
|
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 433,660,774
RAC: 817,079
|
AMD has problems with primegrid and Boinc (until Michael can fix both the server and client ends). And NVIDIA offers CUDA to use with GFN Manual Sieve (until JimB spits out the OCL app). If I was to buy a card right now, I'd say the 1060 is the preferred choice.
I think I'm very close to solving the ATI problem. And by "close" I mean there's more testing to do to ensure that the fix works in all configurations. We've got AP27 working on ATI cards on the test server, and what I did with AP27 should work with GFN. And when the fixes go live, everyone should thank Van -- he spent an inordinate amount of time testing with me. (He's got the hardware configurations needed to do the testing.)
____________
My lucky number is 75898524288+1 |
|
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1957
ID: 352
Credit: 6,144,601,898
RAC: 2,290,195
                        
|
|
So...got 1070 recently (GIGABYTE GeForce GTX 1070 G1 Gaming), hostID 376156.
Testing all GFNs I found that GPU utilization is not what I would have expected, considering there is Skylake i5-6600 feeding the GPU.
I also used Process Tamer to force GFN GPU tasks run in normal priority.
(there are 3 other CPU GFN20 tasks, leaving one core idle)
GFN secs GPU utilization
15 57 73
16 192 85
17l 415 82
18m 625 86
What I found interesting is when testing standalone GFN vs BOINC one.
I got a bit higher GPU usable and GPU core clock went from 1594MHz in BOINC to 1911 or 1936MHz in standalone.
GFN15 73 vs 77 GPU usage, 57 vs 43 secs
GFN16 85 vs 87 GPU usage, 192 vs 165 sec.
BOINC GFN18, 19 (others not tested yet) brings GPU clock to 1911 as well.
EDIT: So, re-running GFN15 under BOINC makes GPU run at 1936Mhz.
Was there some burn-in needed?
>geneferocl_windows.exe -q "30350544^32768+1"
geneferocl 3.3.1-1 (Windows/OpenCL/32-bit)
...
Running on platform 'NVIDIA CUDA', device 'GeForce GTX 1070', version 'OpenCL 1.2 CUDA' and driver '368.81'.
Testing 30350544^32768+1...
Using OCL4 transform
30350544^32768+1 is composite. (RES=a6fae85560e817f6) (245176 digits) (err = 0.0000) (time = 0:00:43) 20:56:36
____________
My stats |
|
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1957
ID: 352
Credit: 6,144,601,898
RAC: 2,290,195
|
|
How is GTX 1070 with larger GFN? It is showing it's full power yet?
With GFN20, GPU usage was about 97-98 percent and was limited by "reliability voltage" as GPU-Z states.
Now with GFN21, the card is hitting TDP limit and running around 1900MHz.
GPU temp 69C, fans at 58%.
(I don't know ambient room temperature as of now since running remotely; probably around 25C but can be up to 29C).
____________
My stats |
|
|
Yves GallotVolunteer developer
Project scientist
Send message
Joined: 19 Aug 12
Posts: 820
ID: 164101
Credit: 305,989,513
RAC: 2,326
|
What I found interesting is when testing standalone GFN vs BOINC one.
I got a bit higher GPU usable and GPU core clock went from 1594MHz in BOINC to 1911 or 1936MHz in standalone.
Is NVIDIA driver "Power Management Mode" set to "Prefer Maximum Performance"?
A new "Optimal Power" mode is available and set as the default. This new Pascal option is an undocumented feature.
|
|
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1957
ID: 352
Credit: 6,144,601,898
RAC: 2,290,195
|
Is NVIDIA driver "Power Management Mode" set to "Prefer Maximum Performance"?
A new "Optimal Power" mode is available and set as the default. This new Pascal option is an undocumented feature.
Good suggestion. It was in Optimal Power, I set it to Maximum preformace.
It doesn't make difference, it is performance capped by reliability voltage and/or total power limit.
____________
My stats |
|
|
|
|
|
Here are the run times for my card:
ASUS GTX 1080 Strix (stock)
Genefer / Runtime / GPU Usage / GPU Temperature / CPU Usage
GFN 15..............51 sec / 66% / 49C / 8%
GFN 16............149 sec / 85% / 51C / 3%
GFN 17L...........305 sec / 82% / 57C / 5%
GFN 17M..........520 sec / 83% / 56C / 1%
GFN 18............965 sec / 85% / 61C / .1%
GFN 19.........3,830 sec / 92% / 62C / .02%
GFN 20.......14,100 sec / 96% / 60C / .02%
GFN 21.......41,935 sec / 98% / 60C / 0%
GFN 22.....168,640 sec / 99% / 60C / 0%
PPS Sieve......318 sec / 70-99% / 60C / 4%
GPU usage for GFN 15 is only 66% and for PPS it's up and down like a yo-yo.
How can I get them close to 100%?
Also with the one GFN 22 task I tested, I had to stop running it several times to do other things on my PC.
For check pointing does it roll back a lot? Would it be faster if I let it run uninterrupted?
Thanks
|
|
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 433,660,774
RAC: 817,079
|
For check pointing does it roll back a lot? Would it be faster if I let it run uninterrupted?
You set the frequency of the checkpointing in BOINC. The app itself will try to checkpoint periodically. On a GPU as fast as yours, it's probably going to attempt to checkpoint every couple of minutes. I say "attempt", because if you have configured BOINC to do checkpoints less frequently, the app will obey your BOINC settings. So if you have BOINC configured to checkpoint once an hour, it will only checkpoint once an hour.
By starting and stopping, you force a GPU task to restart from the last checkpoint, so you'll lose a few minutes of computing time. On tasks as long as those, unless you're starting and stopping a lot, it's not going to make a big difference.
____________
My lucky number is 75898524288+1 |
|
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1957
ID: 352
Credit: 6,144,601,898
RAC: 2,290,195
|
GFN 22.....168,640 sec / 99% / 60C / 0%
There is a GFN22 task on GTX 1080 doing it in 149k secs in this thread.
Getting 183,350 secs on GTX 1070, two GFN22 tasks so far with very similar duration.
____________
My stats |
|
|
mackerelVolunteer tester
Send message
Joined: 2 Oct 08
Posts: 2645
ID: 29980
Credit: 568,565,361
RAC: 198
|
|
Instead of stopping, is it better to suspend in ram? I don't run GPU units often but I do that all the time on CPU ones.
BTW I expect to get a 1070 tomorrow... :) |
|
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 433,660,774
RAC: 817,079
|
Instead of stopping, is it better to suspend in ram? I don't run GPU units often but I do that all the time on CPU ones.
BTW I expect to get a 1070 tomorrow... :)
The "leave tasks in memory" setting only affects CPU tasks. GPU tasks are ALWAYS removed from memory when they're suspended.
____________
My lucky number is 75898524288+1 |
|
|
|
|
Instead of stopping, is it better to suspend in ram? I don't run GPU units often but I do that all the time on CPU ones.
BTW I expect to get a 1070 tomorrow... :)
The "leave tasks in memory" setting only affects CPU tasks. GPU tasks are ALWAYS removed from memory when they're suspended.
Is it a good idea for the CPU tasks? I have it on because intuitively I guessed that it would make it faster if they get suspended and restarted once the CPU usage gets too high, but I'd like to know if it actually helps at all, or even makes things worse. |
|
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14011
ID: 53948
Credit: 433,660,774
RAC: 817,079
|
Instead of stopping, is it better to suspend in ram? I don't run GPU units often but I do that all the time on CPU ones.
BTW I expect to get a 1070 tomorrow... :)
The "leave tasks in memory" setting only affects CPU tasks. GPU tasks are ALWAYS removed from memory when they're suspended.
Is it a good idea for the CPU tasks? I have it on because intuitively I guessed that it would make it faster if they get suspended and restarted once the CPU usage gets too high, but I'd like to know if it actually helps at all, or even makes things worse.
I always leave it on. If it's off, and the task gets suspended, the task must restart from the checkpoint file. All crunching done after the last checkpoint is therefore lost. If the task is kept in memory, it continues from the exact spot where it left off, and nothing is lost.
____________
My lucky number is 75898524288+1 |
|
|
mikey
Send message
Joined: 17 Mar 09
Posts: 1784
ID: 37043
Credit: 791,865,002
RAC: 1,251,156
        
|
Here are the run times for my card:
ASUS GTX 1080 Strix (stock)
Genefer / Runtime / GPU Usage / GPU Temperature / CPU Usage
GFN 15..............51 sec / 66% / 49C / 8%
GFN 16............149 sec / 85% / 51C / 3%
GFN 17L...........305 sec / 82% / 57C / 5%
GFN 17M..........520 sec / 83% / 56C / 1%
GFN 18............965 sec / 85% / 61C / .1%
GFN 19.........3,830 sec / 92% / 62C / .02%
GFN 20.......14,100 sec / 96% / 60C / .02%
GFN 21.......41,935 sec / 98% / 60C / 0%
GFN 22.....168,640 sec / 99% / 60C / 0%
PPS Sieve......318 sec / 70-99% / 60C / 4%
GPU usage for GFN 15 is only 66% and for PPS it's up and down like a yo-yo.
How can I get them close to 100%?
Thanks
Generally speaking you run more than one unit at a time, the first unit uses alot of the gpu but other units running at the same time don't use as much. It's alot of trial and error though, try running 2 units at once and see if you can get into the upper 80's to low 90's, running at 100% means you are maxing out your gpu and that could be bad as you don't know if you are actually slowing things down because of it. So start with running 2 units at once, then after you have some consistent numbers try running 3 units at once, then 4 etc, etc unit you either get into the upper 80's to low 90's for gpu usage OR your workunits crunching times get so long that you are actually crunching SLOWER. Yes each unit will crunch slower when you run more than one unit at a time, but the key is if running one unit at a time takes 60 seconds for each unit to finish, but running 2 units at once only takes 90 seconds then you are ahead of the game. If running 3 units at once takes 2:30 you are STILL ahead of the game as opposed to only running 1 unit at a time.
The OTHER thing to consider is invalid units, if you start coming up with those it's ALSO time to back off and stop increasing the gpu load as you are then losing ground, not gaining any. So even if your gpu CAN do 4 workunits at once in say 3:30, in my example of running one workunit in 60 seconds, but 2 of those workunits won't validate you are not gaining anything and instead LOSING ground!! |
|
|
|
|
Generally speaking you run more than one unit at a time,
I've been here long enough that I should know how to run 2 at once... can you please remind me!
|
|
|
|
|
|
Do I have to run an app_config.xml ?
Can someone post an example for GFN and PPS, Thanks |
|
|
288larssonVolunteer tester
Send message
Joined: 17 Apr 10
Posts: 136
ID: 58815
Credit: 5,609,503,806
RAC: 3,193,977
                      
|
Do I have to run an app_config.xml ?
Can someone post an example for GFN and PPS, Thanks
Hi yes
<app_config>
<app>
<name>genefer15</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.5</cpu_usage>
</gpu_versions>
<fraction_done_exact/>
</app>
<app>
<name>pps_sr2sieve</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.5</cpu_usage>
</gpu_versions>
<fraction_done_exact/>
</app>
</app_config>
it is not faster for me, but you can try |
|
|
|
|
|
Thanks
I have tried running GFN 15, two, three and four at once...
No real difference, maybe 1 or 2 seconds, GPU is still at 66%, the tasks just take longer.
Why can't GFN 15 run at close to 100% GPU?
|
|
|
Yves GallotVolunteer developer
Project scientist
Send message
Joined: 19 Aug 12
Posts: 820
ID: 164101
Credit: 305,989,513
RAC: 2,326
|
Why can't GFN 15 run at close to 100% GPU?
The ratio N=32768 / #cores=2560 is too small to hide GPU instruction latency.
GPU throughput is one instruction per core but pipeline latency is 20 cycles or more (Pascal clock is fast then latency should be high... 30 cycles, more?)
If latency is 30 cycles, 30 * 2560 = 76800 independent operations should be available to hide pipeline latencies.
With N=32768, this is not always the case.
|
|
|
|
|
|
Well running two PPS Sieve at once is far better, GPU is at 100%.
1 task = 318 sec
2 tasks = 518 sec or 259 sec each
Running two at once is about 23% faster for me.
|
|
|
Yves GallotVolunteer developer
Project scientist
Send message
Joined: 19 Aug 12
Posts: 820
ID: 164101
Credit: 305,989,513
RAC: 2,326
|
I have tried running GFN 15, two, three and four at once...
No real difference, maybe 1 or 2 seconds, GPU is still at 66%, the tasks just take longer.
Genefer uses OpenCL local memory (shared data across all work items of a workunit).
Only a fixed and small amount of local memory (48 kB) is available per multi-processor.
Local memory is the bottleneck if you run 2 instances.
|
|
|
