APU memory allocation for GPU?

Message boards : Number crunching : APU memory allocation for GPU?
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile Bill Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 30 Nov 05
Posts: 282
Credit: 6,916,194
RAC: 60
United States
Message 2009075 - Posted: 24 Aug 2019, 12:54:09 UTC

Just curious to see if anyone has changed settings on their APU to use more of the system memory for crunching performance. My 2200G by default has 2.0 GB allocated for the GPU. I'm using between 1.5 and 1.7 GB of that, and I can't help but wonder if that is slowing GPU performance. Currently using 70%/90% of memory, but I assume that is for system memory, not for GPU (whether discrete or not) memory? I have 16 GB total, so no issue with allocating a couple more GB. I'll probably play around with it after WOW is over.
Seti@home classic: 1,456 results, 1.613 years CPU time
ID: 2009075 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5126
Credit: 276,046,078
RAC: 462
Message 2009081 - Posted: 24 Aug 2019, 13:17:45 UTC - in response to Message 2009075.  

Just curious to see if anyone has changed settings on their APU to use more of the system memory for crunching performance. My 2200G by default has 2.0 GB allocated for the GPU. I'm using between 1.5 and 1.7 GB of that, and I can't help but wonder if that is slowing GPU performance. Currently using 70%/90% of memory, but I assume that is for system memory, not for GPU (whether discrete or not) memory? I have 16 GB total, so no issue with allocating a couple more GB. I'll probably play around with it after WOW is over.


When I tried less than 2GB the windows taskmanager started reporting it was using "main memory" too. I don't remember if I ran it long enough to see if it changed the performance.

Tom
A proud member of the OFA (Old Farts Association).
ID: 2009081 · Report as offensive
Profile Kissagogo27 Special Project $75 donor
Avatar

Send message
Joined: 6 Nov 99
Posts: 716
Credit: 8,032,827
RAC: 62
France
Message 2009168 - Posted: 25 Aug 2019, 9:36:11 UTC

after taking a look at some stderr.txt you could modify your command line options

with -period_iterations_num 4 you could add -tt 1500 to speed up your FFT routine here


Fftlength=8,pass=3:Tune: sum=16775.1(ms); min=0.00288(ms); max=505.4(ms); mean=78.76(ms); s_mean=84.41; sleep=75(ms); delta=5; N=213; usual
Fftlength=8,pass=4:Tune: sum=8758.26(ms); min=0.00288(ms); max=283.4(ms); mean=71.79(ms); s_mean=46.18; sleep=45(ms); delta=7; N=122; usual
Fftlength=8,pass=5:Tune: sum=4890.09(ms); min=0.00288(ms); max=163.4(ms); mean=65.2(ms); s_mean=56.68; sleep=45(ms); delta=202; N=75; usual
Fftlength=16,pass=3:Tune: sum=8655.57(ms); min=2.107(ms); max=118.8(ms); mean=61.39(ms); s_mean=51.86; sleep=45(ms); delta=211; N=141; usual
Fftlength=16,pass=4:Tune: sum=5605.09(ms); min=2.838(ms); max=80.53(ms); mean=60.92(ms); s_mean=59.06; sleep=60(ms); delta=276; N=92; usual
Fftlength=16,pass=5:Tune: sum=4018.42(ms); min=2.92(ms); max=83.98(ms); mean=59.09(ms); s_mean=52.64; sleep=45(ms); delta=357; N=68; usual
Fftlength=32,pass=3:Tune: sum=8173.22(ms); min=0.923(ms); max=102.7(ms); mean=59.66(ms); s_mean=56.81; sleep=45(ms); delta=281; N=137; usual
Fftlength=32,pass=4:Tune: sum=5486.69(ms); min=2.882(ms); max=82.95(ms); mean=54.32(ms); s_mean=54.97; sleep=45(ms); delta=338; N=101; usual
Fftlength=32,pass=5:Tune: sum=4079.76(ms); min=0.6144(ms); max=95.38(ms); mean=47.44(ms); s_mean=46.6; sleep=45(ms); delta=475; N=86; usual
Fftlength=64,pass=3:Tune: sum=8425.46(ms); min=0.4523(ms); max=103.1(ms); mean=52.01(ms); s_mean=44.23; sleep=45(ms); delta=305; N=162; usual
Fftlength=64,pass=4:Tune: sum=5814.48(ms); min=0.444(ms); max=68.55(ms); mean=46.52(ms); s_mean=60.01; sleep=60(ms); delta=309; N=125; usual
Fftlength=64,pass=5:Tune: sum=4343.75(ms); min=0.9982(ms); max=51.84(ms); mean=35.9(ms); s_mean=44.43; sleep=45(ms); delta=542; N=121; usual
Fftlength=128,pass=3:Tune: sum=20329.7(ms); min=92.53(ms); max=120.1(ms); mean=107.6(ms); s_mean=104.8; sleep=105(ms); delta=1; N=189; usual
Fftlength=256,pass=3:Tune: sum=24707.1(ms); min=30.16(ms); max=75.43(ms); mean=65.19(ms); s_mean=65.73; sleep=60(ms); delta=1; N=379; usual
Fftlength=512,pass=3:Tune: sum=26651.3(ms); min=15.65(ms); max=65.99(ms); mean=35.11(ms); s_mean=40.29; sleep=30(ms); delta=1; N=759; usual
Fftlength=1024,pass=3:Tune: sum=8220.21(ms); min=5.045(ms); max=6.86(ms); mean=5.419(ms); s_mean=5.576; sleep=0(ms); delta=1; N=1517; usual
Fftlength=2048,pass=3:Tune: sum=5202.03(ms); min=1.399(ms); max=3.792(ms); mean=1.715(ms); s_mean=1.782; sleep=0(ms); delta=1; N=3033; usual


at FFTlength 8 to 1024 u must see some improvement with only one line of pass=3 and high_perf if not u have to set a higher number at -tt


Fftlength=8,pass=3:Tune: sum=16775.1(ms); min=0.00288(ms); max=505.4(ms); mean=78.76(ms); s_mean=84.41; sleep=75(ms); delta=5; N=213; usual => 213 * 78.76 ~= sum time
Fftlength=8,pass=4:Tune: sum=8758.26(ms); min=0.00288(ms); max=283.4(ms); mean=71.79(ms); s_mean=46.18; sleep=45(ms); delta=7; N=122; usual => 122 * 71.79 ~= 8758,38 ~ sum time
Fftlength=8,pass=5:Tune: sum=4890.09(ms); min=0.00288(ms); max=163.4(ms); mean=65.2(ms); s_mean=56.68; sleep=45(ms); delta=202; N=75; usual => 75 * 65.2 ~= sum time


all Fftlength=8 are done in 16775 +8758+ 4890 = 30sec 213 in 3 pass ( best opt) 122 in 4 pass and 75 in 5 pass

the -tt 1500 give more time to do all Fftlength=8 in only 3 pass .

to see the long therm improvement, just take a look at your app details

https://setiathome.berkeley.edu/host_app_versions.php?hostid=8640304

SETI@home v8 8.22 windows_intelx86 (opencl_ati5_SoG_nocal)
Taux de calcul moyen 46.96 GFLOPS

SETI@home v8 8.22 windows_intelx86 (opencl_ati_nocal)
Taux de calcul moyen 42.77 GFLOPS

the estimated Gflops will become better after ^^
ID: 2009168 · Report as offensive
Profile Bill Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 30 Nov 05
Posts: 282
Credit: 6,916,194
RAC: 60
United States
Message 2010159 - Posted: 31 Aug 2019, 15:45:41 UTC

So after looking at the UMA buffer size in my bios, the maximum it can be set to is 2 GB, so I guess I'm stuck with that.
Seti@home classic: 1,456 results, 1.613 years CPU time
ID: 2010159 · Report as offensive
Profile Bill Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 30 Nov 05
Posts: 282
Credit: 6,916,194
RAC: 60
United States
Message 2011270 - Posted: 8 Sep 2019, 17:40:42 UTC

Correction! There was another way in the bios to set the APU dedicated GPU memory. I found a way to increase it to 3 GB. It can go higher, but there is some problem either with the bios or how I'm changing settings that is preventing me from going higher. I'll have to work that out with ASRock. Regardless, I now have 3 GB, and at times I am using 2.6 GB. Of course, in the interim I have changed S@H to run 2 GPU tasks on the APU at a time, so I am sure that is part of the reason. I am probably changing too many settings at once to know what is actually improving performance, so we'll have to see eventually. I just wanted to report memory allocation can be increased.
Seti@home classic: 1,456 results, 1.613 years CPU time
ID: 2011270 · Report as offensive
CryptokiD
Avatar

Send message
Joined: 2 Dec 00
Posts: 150
Credit: 3,216,632
RAC: 0
United States
Message 2011271 - Posted: 8 Sep 2019, 18:15:11 UTC

Everyones computer is different but I have played around with this setting and tried everything from 64 to 2048 and on my system at least there is no difference using anything higher then 256 for my apu's gpu so long as I run up to a max of 1 multibeam and 1 astropulse task at a time.
ID: 2011271 · Report as offensive

Message boards : Number crunching : APU memory allocation for GPU?


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.