Posts by Mad Matt
log in
1) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 18514)
Posted 1437 days ago by Profile Mad Matt
There is no tool to monitor GPU load on OSX. Does anyone have suggested settings for the intel HD 4000? According to wiki, it has "16 execution units and additional features,[b] up to 332.8 GFLOPS at 1300 MHz".


Hi Z,

I for one did not find a lot of improvements when applying the recommended set of instructions for dedicated GPUs on intel_gpus, based on Win OSs here, so mind the gap if there is one.

What seems important if you want to get full output is one core idle on the CPU. I noted after reattaching to the project by going stock (0.528 Core atm) work slowed down by roughly 50% or more. If you just want to run this alongside other stuff, this may be fine, for better output one idle CPU seems crucial.

Matt
2) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 18152)
Posted 1495 days ago by Profile Mad Matt
Never mind got it going.


Good to know and good to see you back, EmSti! :)

Aside of that, I would like to spend a word of thanks to Zydor for his guidance and to Jon for constant GPU pioneering. Collatz was my first ever GPU project, way back on a tiny NV lightyears away from even a 5870, and my first ever Intel GPU project. I think over the years all of my GPUs at least once did Collatz, and there always was an app for them, most of the time an optimized one.

Cheers,
Matt
3) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 17600)
Posted 1570 days ago by Profile Mad Matt
Just checked the OEM support website, they offer a general Catalyst v13.8 mobility beta package from end of August and a notebock-specific driver package from May 2013.

I am not sure if I should go the driver upgrade path yet. This rig is crunching happily at POEM atm with a BOINC RAC of approx. 30k and the games I play from time to time are running fine, too...


I don't know how OpenCL apps are compiled, but the first POEM OpenCL app was written more than one year before Collatz OpenCL, if I remember correctly. So there may be a link to driver compatibility. Even more so since the 'never-settle' update was during that period enhancing AMDs strongly.

If you were using a desktop, I'd say definitely try as you also may see better performance on other projects. With notebooks this may look differently, but then again you should have backups and tech-support...

Just my guess, maybe it may help.
4) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 17586)
Posted 1573 days ago by Profile Mad Matt
Suggestions are welcome but this is first of all FYI.


Not knowing a thing about mobile GPUs and their architecture at all, did you monitor your temps during crunching and do you have better/normal results on other projects to compare this performance?


5) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 17569)
Posted 1575 days ago by Profile Mad Matt
Regarding clocks, at least without adapting power many 7970 GPUs will not go beyond 1125 without failing, e.g. my XFX 7970.


Matt

Just realised .... sorry, I was slow off the start gate, only just twigged....

Are you using MSI Afterburner? On the face of it, it seems you are not, as Afterburner will give the ability to go beyond the AMD imposed limit of 1125 etc.



Yes, I am using it. I won't say you cannot run more, but in contrast to the Lightnings coming with a hardcoded high-boost, 'normal' Tahiti xt chips seem to fail beyond 1125 MHz without increasing powerplay. Likely you may see higher clocks when adapting powerplay to higher values, but I did not want to take the cards that far.

Cheers for the help
Matt
6) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 17526)
Posted 1582 days ago by Profile Mad Matt
Your Welcome ....

Just had a peek at your results so far. You are probably getting close to max re GPU speed looking at the timings which are pretty well the same as mine as such. Mine is with a AMD Phenom(tm) II X6 1090T (my main machine is still down, should be back on next week hopefully), so you should be able to go a little further with your better CPU, probably down to an average of around 525-540 secs.

Don't push to these last steps overnight, do it during the day when you can spot the "one step too far" crash - as this would be the last "downward" step its likely you'll hit the Wall at some point, so best done during the day when you can quickly spot an issue.


Adding to that:

verbose=1
items_per_kernel=18 start -->22 best possible
kernels_per_reduction=7 -->9 best possible
threads=8
sleep=1
build_options=-Werror

These best settings do not seem to pose problems to any 7970 or an MSI GTX 680 Lightning. Unless of course lag is a problem. ;) Regarding clocks, at least without adapting power many 7970 GPUs will not go beyond 1125 without failing, e.g. my XFX 7970.

I kept RAM to 1500 throughout all 7970s as I found it more rewarding to pay attention to the CPU clocks which seemed to cause more of a difference. So if you don't toast the CPU on its CPU-only tasks, investing some of the thermal headroom may be well spent on clocking CPU and CPU-RAM first.
7) Message boards : Number crunching : Optimizing Collatz v6.xx OpenCL and CUDA Applications (Message 17511)
Posted 1584 days ago by Profile Mad Matt
Solo_Collatz Useage

Afterburner Setting:
Core Voltage: 1213
Power Limit: +0
Core Clock: 1225
Memory Clock: 1675
Fan Speed: 55% (running temperature of the GPU is 74 degrees, utilisation is 98%)

The XML Configuration File I used is the same as the one Slicker posted ie:
<configuration>
verbose=1
items_per_kernel=20
kernels_per_reduction=9
threads=8
sleep=1
build_options=-Werror
</configuration>

My app_config is:
<app_config>
<app>
<name>solo_collatz</name>
<max_concurrent>1</max_concurrent>
<gpu_versions>
<gpu_usage>1.0</gpu_usage>
<cpu_usage>1.0</cpu_usage>
</gpu_versions>
</app>
</app_config>




First off - standard disclaimer ..... this pushes the card beyond standard use, and into territory not supported by either NVidia or AMD - you do so at your own risk - don't yell at me if you burn out the card :)

(its unlikely ..... but ..... its non standard use, so its at your risk, not mine)

The key is optimising the output. To do that you must paste this ......


verbose=1
items_per_kernel=18
kernels_per_reduction=7
threads=8
sleep=1
build_options=-Werror


.... into the existing file "solo_collatz_4.07_windows_x86_64__opencl_ati_100" marked as a CONFIG file. Probably the directory on yours is

C:\ProgramData\BOINC\projects\boinc.thesonntags.com_collatz

To do so, open that file with notepad, don't mess with the file extension, just paste the lines shown above into it (do not use the "<configuration>" elements in my post above, just paste the lines of text.

When pasted, save the file - don't mess with extensions etc - just save as is, "default". You will need to restart BOINC, and then the app for it to take effect. I have set cautious values of 18 & 7 in the lines above - in reality I am running them as 22 and 9 - however take it slow and steady at first, get it running, then increase each value separately in stages. If you reach the max of 22 & 9 - well, fine - but you may crash, if the latter back off one step with those values.



Hammer, nail. Thank you Zydor. Went to your values 22 and 9 directly as it seems I am using the same GPU. Right away down from >1500 to <600 seconds. Now curious to see how it unfolds. Still at 1150/1500 clocks so there seems room for improvement.

Cheers,
Matt
8) Message boards : News : Collatz v3.17 CUDA Testers for Windows Needed (Message 15948)
Posted 1820 days ago by Profile Mad Matt
Two hosts running 3.17, BOINC 7.0.28. Single GPUs. Kept all parameters on default. So far no problems, but just one WU validated yet:

http://boinc.thesonntags.com/collatz/show_host_detail.php?hostid=120719
http://boinc.thesonntags.com/collatz/show_host_detail.php?hostid=120717
9) Message boards : Number crunching : Combining extremely different (CUDA) GPUs? (Message 8374)
Posted 2777 days ago by Profile Mad Matt
The only way I know of to do this would be to set up a VM with a separate BOINC install and use the various BOINC settings (e.g., use CPU, use GPU, allocate % of CPU, etc.) to get the mix desired.



Not sure if VMware managed to do this, but I always thought virtual machines use emulated video cards?
10) Message boards : Number crunching : Combining extremely different (CUDA) GPUs? (Message 8317)
Posted 2782 days ago by Profile Mad Matt

My first inclination was to say no, just like you cannot on a multi-core cpu. BUT I then thought of using an app_info file and using the gpu 0 or gpu 1 setting to designate whether the gpu is active or not for each project and thought maybe. You will have to have an expert answer this but it could be interesting.


Does sound like a plan, but yes, expert advice is needed here.

Practically it should work since today I have seen two Collatz and one GPUGRID crunched at the same time due to a drift in crunching times. But I am also wondering how this may affect computing speed. I think the 8600 GT did have some fantastic results I did not see yet, but then again, it could slow down the GTX 295.

Does anyone know if CUDA processes run by the same driver really are discrete?
11) Message boards : Number crunching : Combining extremely different (CUDA) GPUs? (Message 8303)
Posted 2783 days ago by Profile Mad Matt
I am wondering if it could make any sense to combine extremely different GPUs like a GTX295 and an old 8600GT on one system. Of course the old card would not be able to run every project and the way BOINC seems to handle GPU-WUs (first in first out) for single GPUs seems to hinder this from working - in theory.

Will BOINC be able so separate between the GPUs, so I can assign them to two different projects at times, whenever I would like to run e.g. GPUGRID or Milkyway on the GTX 295?

Sorry if this has been asked before, but I could not find the answer yet, neither here nor at GPUGRID.
12) Message boards : Number crunching : CUDA 2.04 Application Hints, Tips, Performance, and Issues (Message 8299)
Posted 2783 days ago by Profile Mad Matt

I assume if running on highest GPU setting, the daily output should really increase heavily based on this. However, I will not use this very often since most of my favourite projects still are CPU-intensive.


Strange. Optimization only works in one way by reducing CPU overall usage massively. I assumed to see a huge boot with <cmdline>L13 I8 S0</cmdline>, but in fact only CPU usage was boosted extremely while total crunching time did not seem to change at all. So, back to <cmdline>L13 I8 S1</cmdline> here.

This happened both with the 8600 GT and 9800 GT.
13) Message boards : Number crunching : CUDA 2.04 Application Hints, Tips, Performance, and Issues (Message 8245)
Posted 2787 days ago by Profile Mad Matt
For both of my cards, 9800 GT (W7-64) and 8600 GT (XP-64), I did not notice any remarkable change in crunching time when using the setting "To run as fast as possible on both CPU and GPU". However, when switching to "run as fast as possible (max GPU WUs per day)" CPU effectiveness went up to about 0.15 in Boincview, while the balanced settings drastically reduced CPU usage compared to stock applications.

I assume if running on highest GPU setting, the daily output should really increase heavily based on this. However, I will not use this very often since most of my favourite projects still are CPU-intensive.
14) Message boards : News : Windows CUDA 2.04 Opt App Now Available (Message 8243)
Posted 2787 days ago by Profile Mad Matt
Obsolete question by now. ;) It does run properly on 64 bit Windows XP and 7.

So, next question... could you gain even more by optimizing for 64 bit or creating a native 64 bit application?
15) Message boards : News : Windows CUDA 2.04 Opt App Now Available (Message 8236)
Posted 2787 days ago by Profile Mad Matt
I suppose it just won't work on 64bit Win?
16) Message boards : Number crunching : CUDA 23 app hangs up on exit/pause (Message 2655)
Posted 3025 days ago by Profile Mad Matt
Does this have something to do with your optimisation for CUDA? Of course I do not have an insight into how many milliseconds other projects need on CUDA, but Seti, Folding or GPUGRID appear to stop immediately or almost immediately if asked to pause or quit.
17) Message boards : Number crunching : CUDA 23 app hangs up on exit/pause (Message 2488)
Posted 3029 days ago by Profile Mad Matt
It's BOINC 6.6.38 and most of the time it's set to use GPU while computer is in use.
18) Message boards : Number crunching : CUDA 23 app hangs up on exit/pause (Message 2482)
Posted 3030 days ago by Profile Mad Matt
I should add that after almost every pause/resume of BOINC I do have two instances of Collatz running. But both will not progress until they are forced to shut down and BOINC is restarted.

Any idea for this strange behaviour?
Thanks
19) Message boards : Number crunching : CUDA 23 app hangs up on exit/pause (Message 2451)
Posted 3031 days ago by Profile Mad Matt
Everything is running fine for me, but frequently on exit or pause commands from BOINC manager the Collatz 2.03 CUDA23 app hangs up and seems to keep running on the card.

Something like this will happen:

04.10.2009 01:20:59 Collatz Conjecture Task collatz_1254320457_53046_2 exited with zero status but no 'finished' file
04.10.2009 01:20:59 Collatz Conjecture If this happens repeatedly you may need to reset the project.
04.10.2009 01:20:59 Collatz Conjecture Restarting task collatz_1254320457_53046_2 using collatz version 203
04.10.2009 01:21:40 Collatz Conjecture Task collatz_1254320457_53046_2 exited with zero status but no 'finished' file
04.10.2009 01:21:40 Collatz Conjecture If this happens repeatedly you may need to reset the project.
04.10.2009 01:21:40 Collatz Conjecture Restarting task collatz_1254320457_53046_2 using collatz version 203
04.10.2009 01:22:21 Collatz Conjecture Task collatz_1254320457_53046_2 exited with zero status but no 'finished' file
04.10.2009 01:22:21 Collatz Conjecture If this happens repeatedly you may need to reset the project.
04.10.2009 01:22:21 Collatz Conjecture Restarting task collatz_1254320457_53046_2 using collatz version 203
04.10.2009 01:23:02 Collatz Conjecture Task collatz_1254320457_53046_2 exited with zero status but no 'finished' file
04.10.2009 01:23:02 Collatz Conjecture If this happens repeatedly you may need to reset the project.
04.10.2009 01:23:02 Collatz Conjecture Restarting task collatz_1254320457_53046_2 using collatz version 203
04.10.2009 01:23:43 Collatz Conjecture Task collatz_1254320457_53046_2 exited with zero status but no 'finished' file
04.10.2009 01:23:43 Collatz Conjecture If this happens repeatedly you may need to reset the project.
04.10.2009 01:23:43 Collatz Conjecture Restarting task collatz_1254320457_53046_2 using collatz version 203
04.10.2009 01:24:24 Collatz Conjecture Task collatz_1254320457_53046_2 exited with zero status but no 'finished' file
04.10.2009 01:24:24 Collatz Conjecture If this happens repeatedly you may need to reset the project.
04.10.2009 01:24:24 Collatz Conjecture Restarting task collatz_1254320457_53046_2 using collatz version 203
04.10.2009 01:25:05 Collatz Conjecture Task collatz_1254320457_53046_2 exited with zero status but no 'finished' file
04.10.2009 01:25:05 Collatz Conjecture If this happens repeatedly you may need to reset the project.
04.10.2009 01:25:05 Collatz Conjecture Restarting task collatz_1254320457_53046_2 using collatz version 203

After closing the app in task manager and restarting BOINC everything is back to normal.




Main page · Your account · Message boards


Copyright © 2018 Jon Sonntag; All rights reserved.