EVGA

Good News over on PG

Author
bcavnaugh
The Crunchinator
  • Total Posts : 37537
  • Reward points : 0
  • Joined: 2012/09/18 17:31:18
  • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
  • Status: offline
  • Ribbons : 272
2019/05/22 13:20:14 (permalink)
AVX-512 Now Supported by LLR
All of PrimeGrid's LLR applications now support AVX-512 on CPUs with that capability.  Those of you that have been using app_info.xml/anonymous platform to run LLR 3.8.23 may now use the stock app if you wish, which is also LLR 3.8.23.
22 May 2019 | 0:22:04 UTC   · Comment
 


#1

14 Replies Related Threads

    bill1024
    CLASSIFIED ULTRA Member
    • Total Posts : 7098
    • Reward points : 0
    • Joined: 2008/10/18 01:01:10
    • Status: offline
    • Ribbons : 43
    Re: Good News over on PG 2019/05/22 13:44:28 (permalink)
    Is that faster than FMA3 ?
     

     Life is too short to carry a cheap pocket knife

     U25ITA93JV   
     
    #2
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 14:22:51 (permalink)
    I have no idea but I am going to run some today. My Offset for AVX3 is now set to 3 @ 4300MHz so it should run @ 4000MHz
    My X299's have AVX512F AVX3 Setting in the Bios. First time we have an app that can use it on a BOINC Project.
    From Michael Goetz 
     
     If you have a Skylake-X CPU, or one of the server equivalents with two AVX-512 units, you could see up to double the performance than before.
    This may be limited e.g. by ram bandwidth, or (lack of) multi-core scaling.
    The CPU will run hotter. Running 6x PPSE my 7800X (delid + liquid metal) at stock clocks was hitting around 100C with a Noctua D14/D15 cooler (not sure exact model).
    I still want to underclock the AVX-512 some time...
    post edited by bcavnaugh - 2019/05/22 14:37:12


    #3
    bill1024
    CLASSIFIED ULTRA Member
    • Total Posts : 7098
    • Reward points : 0
    • Joined: 2008/10/18 01:01:10
    • Status: offline
    • Ribbons : 43
    Re: Good News over on PG 2019/05/22 14:55:09 (permalink)
    Interesting, thanks Bill.  I don't think I have any CPUs with that AVX 512 instruction set.
    I have 5930k, 8600k, 8086k  x99 and z370  chip set, they are my newest, which is not all that old at all. 
     
    If it runs hotter than FMA3 or AVX2 and you have to back it off, is it still faster. Guess we will find out after testing.
    Will the program use AVX 512 or will it use FMA3. Seems to want to use FMA3 not AVX for some reason.

     Life is too short to carry a cheap pocket knife

     U25ITA93JV   
     
    #4
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 14:57:53 (permalink)
    I would hope so or think so.


    #5
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 15:49:18 (permalink)


    #6
    bill1024
    CLASSIFIED ULTRA Member
    • Total Posts : 7098
    • Reward points : 0
    • Joined: 2008/10/18 01:01:10
    • Status: offline
    • Ribbons : 43
    Re: Good News over on PG 2019/05/22 22:05:08 (permalink)
    bcavnaugh
    Running PPS-Mega (LLR) v8.03
    3 @ 4MT
    https://www.primegrid.com/results.php?hostid=920343
    Old School 4MT
    https://www.primegrid.com/results.php?hostid=939093
    Old School 6MT
    https://www.primegrid.com/results.php?hostid=807978
     
     


    Does that mean you are running 3 tasks 4 threads MT on one CPU? What CPU and how much cache?
    What RAM and at what speed?
    What's old school? Looks like it is still using FMA3, no?
     
    I thought I  read that it is best to try to keep the tasks in cache memory and try not to use system mem.
    You're getting more throughput that way? 

     Life is too short to carry a cheap pocket knife

     U25ITA93JV   
     
    #7
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 22:15:32 (permalink)
    bill1024
    bcavnaugh
    Running PPS-Mega (LLR) v8.03
    3 @ 4MT
    https://www.primegrid.com/results.php?hostid=920343
    Old School 4MT
    https://www.primegrid.com/results.php?hostid=939093
    Old School 6MT
    https://www.primegrid.com/results.php?hostid=807978

    Does that mean you are running 3 tasks 4 threads MT on one CPU? What CPU and how much cache?
    What RAM and at what speed? Ripjaws V Series 3200 (PC4 25600) F4-3200C14Q-64GVK
    What's old school? Looks like it is still using FMA3, no? YES
     
    I thought I  read that it is best to try to keep the tasks in cache memory and try not to use system mem.
    You're getting more throughput that way? 

    Yes
    Ripjaws V Series 3200 (PC4 25600) F4-3200C14Q-64GVK
    Intel® Core™ i9-7940X X-series Processor
    # of AVX-512 FMA Units 2 and 19.25M Cache 
    See https://www.primegrid.com/forum_thread.php?id=8601&nowrap=true#129682
    UPDATE May 22nd:  It has come to my attention that while CPUs with 2 AVX-512 execution units gain a substantial boost in performance, mid-range CPUs with only 1 AVX-512 execution unit may see a significant decrease in performance with the new LLR app.  Obviously, this is not intended.  For the time being there is no workaround for this.  If you have a CPU that supports AVX-512, but has only a single AVX-512 execution unit, you may want to use the anonymous platform mechanism (app_info.xml) to run the older version of LLR.  With a challenge starting tomorrow, we won't make changes to the app until at least a week.  We apologize for any inconvenience this may cause.
     
    https://www.primegrid.com/result.php?resultid=1001845684 

    Stderr output

    <core_client_version>7.14.2</core_client_version>
    <![CDATA[
    <stderr_txt>
    BOINC llr wrapper (version 8.00)
    Using Jean Penne's llr (64 bit)
    LLR Program - Version 3.8.23, using Gwnum Library Version 29.8

    LLR command line: primegrid_cllr.exe -d -oDiskWriteTime=1 llr.in
    Using all-complex AVX-512 FFT length 240K, Pass1=640, Pass2=384, clm=1, a = 5
    16:46:55 (5792): called boinc_finish

    </stderr_txt>
    ]]>

    https://www.primegrid.com/result.php?resultid=1001840622

    Stderr output

    <core_client_version>7.14.2</core_client_version>
    <![CDATA[
    <stderr_txt>
    BOINC llr wrapper (version 8.00)
    Using Jean Penne's llr (64 bit)
    LLR Program - Version 3.8.23, using Gwnum Library Version 29.8

    LLR command line: primegrid_cllr.exe -d -oDiskWriteTime=1 -oThreadsPerTest=4 llr.in
    Using all-complex AVX-512 FFT length 280K, Pass1=640, Pass2=448, clm=1, 4 threads, a = 3
    16:16:36 (3520): called boinc_finish

    </stderr_txt>
    ]]>

     
    Old School AMD 6MT
    https://www.primegrid.com/result.php?resultid=1001829364

    Stderr output

    <core_client_version>7.14.2</core_client_version>
    <![CDATA[
    <stderr_txt>
    BOINC llr wrapper (version 8.00)
    Using Jean Penne's llr (64 bit)
    LLR Program - Version 3.8.23, using Gwnum Library Version 29.8

    LLR command line: primegrid_cllr.exe -d -oDiskWriteTime=1 -oThreadsPerTest=6 llr.in
    Using all-complex FMA3 FFT length 240K, Pass1=1280, Pass2=192, clm=2, 6 threads, a = 3
    16:46:05 (7048): called boinc_finish

    </stderr_txt>
    ]]>

     
    Non-MT
    https://www.primegrid.com/result.php?resultid=1001850152

    Stderr output

    <core_client_version>7.14.2</core_client_version>
    <![CDATA[
    <stderr_txt>
    BOINC llr wrapper (version 8.00)
    Using Jean Penne's llr (64 bit)
    LLR Program - Version 3.8.23, using Gwnum Library Version 29.8

    LLR command line: primegrid_cllr.exe -d -oDiskWriteTime=1 llr.in
    Using all-complex FMA3 FFT length 240K, Pass1=1280, Pass2=192, clm=2, a = 3
    18:24:55 (4428): called boinc_finish

    </stderr_txt>
    ]]>

     
    Old School Intel 4MT
    https://www.primegrid.com/result.php?resultid=1001842263

    Stderr output

    <core_client_version>7.14.2</core_client_version>
    <![CDATA[
    <stderr_txt>
    BOINC llr wrapper (version 8.00)
    Using Jean Penne's llr (64 bit)
    LLR Program - Version 3.8.23, using Gwnum Library Version 29.8

    LLR command line: primegrid_cllr.exe -d -oDiskWriteTime=1 -oThreadsPerTest=4 llr.in
    Using all-complex FMA3 FFT length 240K, Pass1=1280, Pass2=192, clm=2, 4 threads, a = 7
    16:49:32 (5144): called boinc_finish

    </stderr_txt>
    ]]>

     
    post edited by bcavnaugh - 2019/05/22 22:35:28


    #8
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 22:36:33 (permalink)
    I would have to run one task to see if I can keep it in the 19.25M Cache


    #9
    bill1024
    CLASSIFIED ULTRA Member
    • Total Posts : 7098
    • Reward points : 0
    • Joined: 2008/10/18 01:01:10
    • Status: offline
    • Ribbons : 43
    Re: Good News over on PG 2019/05/22 22:38:00 (permalink)
    Very good Bill, Nice CPU by the way!

     Life is too short to carry a cheap pocket knife

     U25ITA93JV   
     
    #10
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 22:39:47 (permalink)
    I will run 2 over the next Event, and thank you.
    + 1 Core i9-7900X
    post edited by bcavnaugh - 2019/05/22 22:41:26


    #11
    bill1024
    CLASSIFIED ULTRA Member
    • Total Posts : 7098
    • Reward points : 0
    • Joined: 2008/10/18 01:01:10
    • Status: offline
    • Ribbons : 43
    Re: Good News over on PG 2019/05/22 22:40:16 (permalink)
    You have to know the size of the tasks you're runnning. that CPU has a nice chunk of cache.
    Some sub projects are small, like Sophie, others are much bigger.
     

     Life is too short to carry a cheap pocket knife

     U25ITA93JV   
     
    #12
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 22:42:45 (permalink)
    Then the "The Riesel Problem LLR" will be way to large.


    #13
    bill1024
    CLASSIFIED ULTRA Member
    • Total Posts : 7098
    • Reward points : 0
    • Joined: 2008/10/18 01:01:10
    • Status: offline
    • Ribbons : 43
    Re: Good News over on PG 2019/05/22 22:54:31 (permalink)
    It was too big for my 8086, I tried two at a time and my time more than tripled.
    Maybe on my xeons I can, they have bigger cache.
     
    Try one at a time, then try two, if you can do two times the tasks in less that two times the time, worth doing.
     
    Maybe with 14 cores doing 14 at a time may take a long time, but maybe able to do 14 that way faster than 14 fast MT.
    Funny how that works sometimes. But primes is also about getting the 1st task in.
    But this sub project, primes are very rare, I doubt one will be found this challenge, challenges are about throughput. 
    Like with the 4P AMD on some projects they seemed slow, but dropping 48 at a time made up for it. 
    Too late to test every which way now.

     Life is too short to carry a cheap pocket knife

     U25ITA93JV   
     
    #14
    bcavnaugh
    The Crunchinator
    • Total Posts : 37537
    • Reward points : 0
    • Joined: 2012/09/18 17:31:18
    • Location: USA Affiliate E5L3CTGE12 Associate 9E88QK5L7811G3H
    • Status: offline
    • Ribbons : 272
    Re: Good News over on PG 2019/05/22 22:57:22 (permalink)
    Running 14 would be 100% on the CPU and would run way to Hot.
    Maybe 2MT and run 6 of them


    #15
    Jump to:
  • Back to Mobile