7970M / astropulse

<avg_ncpus>0.50</avg_ncpus>
        <max_ncpus>0.50</max_ncpus>
        <platform>windows_x86_64</platform>
        <plan_class>ati13ati</plan_class>
        <cmdline>-instances_per_device 4 -unroll 14 -ffa_block 4096 -ffa_block_fetch 2048 -sbs 128</cmdline>
        <coproc>
            <type>ATI</type>
            <count>0.25</count>

The last section being 7950 / astropulse, of course.

Thank you for the link to the lunatics forum, which helped. Just a short report, how it worked so far: Using the lunatics v0.40 installer and editing the app_info file afterwards I run now two instances on my mobile 7970M (each using one core of a quadcore w/ SMT) and four instances on my desktop 7950 (each using 1/2 core of a quadcore w/ CMT). This runs fine. Did no synthetic tests, however, since I am in the middle of a challenge. Here's what I changed:

7970M / multibeam

        <avg_ncpus>1.00</avg_ncpus>
        <max_ncpus>1.00</max_ncpus>
        <plan_class>ati13ati</plan_class>
        <cmdline>-period_iterations_num 20 -instances_per_device 2</cmdline>
        <coproc>
            <type>ATI</type>
            <count>0.5</count>

7970M / astropulse

        <avg_ncpus>1.00</avg_ncpus>
        <max_ncpus>1.00</max_ncpus>
        <platform>windows_intelx86</platform>
        <plan_class>ati13ati</plan_class>
        <cmdline>-instances_per_device 2 -unroll 10 -ffa_block 4096 -ffa_block_fetch 2048 -sbs 128</cmdline>
        <coproc>
            <type>ATI</type>
            <count>0.5</count>

7950 / multibeam

<avg_ncpus>0.50</avg_ncpus>
        <max_ncpus>0.50</max_ncpus>
        <plan_class>ati13ati</plan_class>
        <cmdline>-period_iterations_num 30 -instances_per_device 4</cmdline>
        <coproc>
            <type>ATI</type>
            <count>0.25</count>

7970M / astropulse

<avg_ncpus>0.50</avg_ncpus>
        <max_ncpus>0.50</max_ncpus>
        <platform>windows_x86_64</platform>
        <plan_class>ati13ati</plan_class>
        <cmdline>-instances_per_device 4 -unroll 14 -ffa_block 4096 -ffa_block_fetch 2048 -sbs 128</cmdline>
        <coproc>
            <type>ATI</type>
            <count>0.25</count>

Best regards

Here ATi performance tuning conducted: http://lunatics.kwsn.net/12-gpu-crunching/new-set-of-test-tasks-for-gpu-ap.msg49310.html#msg49310

AP only:
(...)
-unroll 4 Optimal at half the number of Compute Units of the GPU. Lower values also reduce VRAM use. Decrease if you experience lags.

First of all, thank you for all the work on these applications!

I'm currently trying to optimize options for my 7970M (which is more or less a downclocked desktop 7870). Richard wrote that the optimal unroll value is half of the CUs, in my case 20 / 2 = 10. Question: Is this still true, if two instances run on the card or should it be lower then?

Thanks in advance / Best regards

And thank you for the kind words.

Although the release notes are published under my name, they're actually a collaborative effort, bringing together suggestions and knowledge gathered from a wide group of developers and testers. In particular, I only have personal experience of NVidia cards, so I don't think I ought to try to elaborate on the sections contributed by my ATI colleagues.

Mike or Raistmer are probably best placed to answer your direct question: meanwhile, you might be interested in the adjacent GPU AP performance tuning thread - although they, too, have started with NVidia applications.

AP only:
(...)
-unroll 4 Optimal at half the number of Compute Units of the GPU. Lower values also reduce VRAM use. Decrease if you experience lags.

First of all, thank you for all the work on these applications!

I'm currently trying to optimize options for my 7970M (which is more or less a downclocked desktop 7870). Richard wrote that the optimal unroll value is half of the CUs, in my case 20 / 2 = 10. Question: Is this still true, if two instances run on the card or should it be lower then?

Thanks in advance / Best regards

Do you (anybody) see the screenshots now?
It appears for me that the hosting site (prikachi.com) is not working for the moment.
I can repost them using different hosting site.

To see are 'they' running:
in Windows Task Manager (or Process Explorer, ...) - check if any or all of the following processes run:
AK_v8b2_win_SSE3_AMD.exe
ap_6.01r557_SSE2_331_AVX.exe
MB6_win_x86_SSE3_OpenCL_ATi_HD5_r390.exe
AP6_win_x86_SSE2_OpenCL_ATI_r555.exe

---

 


- ALF - "Find out what you don't do well ..... then don't do it!" :)
 

Ah. I thought that the benchmarks would change because of the additional computing power of the Graphics Processor. In that case, I'll go back, reset the whole thing back to BlBgs settings (big thanks for the screencaps), and give SETI about a week to sort things out.

Once it does, I should start seeing an increase in the User Averages, correct?

The benchmark is just for BOINC to get an idea of how fast your machine is to request work from projects. IIRC one of the two numbers isn't actually used any longer.

Your average will creep up over time. Depending on the speed of the machine and how often it runs. Generally a few weeks, but you can check the daily stat numbers form the stat websites.

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

Ah. I thought that the benchmarks would change because of the additional computing power of the Graphics Processor. In that case, I'll go back, reset the whole thing back to BlBgs settings (big thanks for the screencaps), and give SETI about a week to sort things out.

Once it does, I should start seeing an increase in the User Averages, correct?

---

The installer does not change the benchmarks in any way.

It simply installs replacement SETI science applications, as the name implies.

You will not see any change in the estimated runtime, either, until BOINC has had a chance to assess how well the new applications perform. But - assuming you followed the screenshots that BilBg posted - you could expect that SETI and Astropulse CPU tasks finish in half the time that the stock applications took for equivalent tasks.

Don't change anything yet, give yourself time to see how things are working first.

OK. I got the card installed, and it's driver loaded up. BOINC recognizes that it is there.

http://setiathome.berkeley.edu/show_host_detail.php?hostid=6560195

Everything is running along smoothly, but I am not seeing an increase in benchmark speed. I've been fiddling with allowing astro pulse and other settings, nothing seems to work. Should I just uninstall seti, Reinstall it, and run the lunatics 64 installer and see if that perks it up? If I do, will that lose my progress?

---

These are the settings for using AMD CPU + ATI GPU for SETI
(if you have AMD Athlon(tm) II X4 630 Processor, AMD Radeon HD 6570, Windows 7 x64)






P.S.
- The screenshots are from the 32 bit Lunatics Installer, I don't know if the 64 bit looks the same.
(obviously the 'SSE (p3)' and 'SSE2' will not be in the 64 bit Lunatics Installer as these CPUs are not 64 bit)

- There is no file called lunaticswin6440.04.exe
The file (Lunatics Installer 64 bit) is Lunatics_Win64_v0.40_setup.exe
http://lunatics.kwsn.net/index.php?module=Downloads;catd=9

- The Lunatics Installer does NOT install any driver for the GPU
It just puts some files (.exe, .dll, .xml, ...) in the SETI directory (....\BOINC\projects\setiathome.berkeley.edu\)

---

 


- ALF - "Find out what you don't do well ..... then don't do it!" :)
 

I hate to sound dimwitted, but I just want to be clear. SO when I get the Radeon in (new power source just got in the other day), I run the Lunatics Installer, and select GPU. It installs a driver for the GPU. After that I can select for multibeam and astropulse in the SETI control panel, right?

Nope, the Lunatics Installer doesn't install any drivers, you have to install them yourself prior to running the Lunatics Installer,

ATI Driver downloads

The Lunatics Installer only installs the Optimised apps that you select, ie, a CPU Multibeam app, a CPU Astropulse v6 app, a Cuda MB app, an ATI OpenCL MB app, an ATI OpenCL Astropulse v6 app or a ATI Brook+ Hybrid Astropulse v6 app.

Claggy

I hate to sound dimwitted, but I just want to be clear. SO when I get the Radeon in (new power source just got in the other day), I run the Lunatics Installer, and select GPU. It installs a driver for the GPU. After that I can select for multibeam and astropulse in the SETI control panel, right?

Not quite.

First, you have to install a driver to suit your hardware. The card will come with its own installation CD/DVD (supplied by the card manufacturer), or you can download the current driver from the AMD download page. Make sure you choose the right one for your operating system - if Windows, Win7 or XP, 32-bit or 64-bit. Make sure installation is complete, and all necessary computer reboots have been completed.

Then, run the Lunatics Installer. It will let you choose SETI programs - Astropulse and/or MultiBeam, CPU and/or GPU - as you wish, and copy them to your machine ready for BOINC to use.

After that, manage your crunching via this website and the BOINC Manager as usual. There is no 'SETI control panel' as such.

You mean your Radeon HD6570? Then you should fill in the ATI part of the installer. Or do you want to only run CPU? (Check CPU capabillities like
SS(S)E3(x)4.1/4.2).
Why don't you use your 'spanking new Radeon HD6570', you can run MB
and AstroPulse work which works amazingly effective on an ATI 5000 or higher
series. A runtime of less then 1 hour to a few hours, CPU time is very
depending on % of RADAR-Blanking.

You can run multiple WUs per GPU and crashes are very rare,
AV scans(!) disable them for BOINC also WINDOWs updates are the most common reasons. The rev.390 MB app. and rev.555 app. for AstroPulse are stable and
haven't seen much errors from these. have a look at this
host
an i7-2600 and 2 ATI 5870 GPUs.


I freed up one 'thread', 0.5 core of an i7 and used BOINCs setting:
use 87.50 of CPU, I used your method in my app_info.xml trying
to improve loading of the 2nd ATI 5870 GPU.
But stopped running multiple WUs on a GPU, less effective, easier to compaire
run-times v.s NVidia GPUs. Although GPU load running 1 a GPU are 67% average.
(GPU 0, always has a higher load as GPU 1)![/quote]


I hate to sound dimwitted, but I just want to be clear. SO when I get the Radeon in (new power source just got in the other day), I run the Lunatics Installer, and select GPU. It installs a driver for the GPU. After that I can select for multibeam and astropulse in the SETI control panel, right?

---

How about I just leave things as they are and see if they work better on the new card?

If you "leave things as they are" (= wrong apps) you will generate errors.

Either return to stock/standard apps (Remove/Add project)
or run again the Lunatics Installer (Lunatics_Win64_v0.40_setup.exe)
and choose the proper apps - AMD for CPU and nothing for GPU




P.S.
ap_6.01r557_SSE2_331_AVX.exe work OK and is faster on my AMD Athlon(tm) 64 Processor 3500+
http://setiathome.berkeley.edu/forum_thread.php?id=67407&nowrap=true#1213064

Are those apps. used at Bêta, or already released for upgrade of rev.(330(?)390
for MB and rev.555 for AstroPulse? (rev.557?) for GPU
Ahh, but these are the C.P.U. app.s not A.T.I. (AMD) G.P.U.s !
(Should install those or newer(?) on Bêta, anyway)!?

An i7(i5 & i3) should benefit from AVX also SSE3(or higher).

---

A couple of comments on freeing one core, as a refinement but not as a recommendation.

For any host with up to 100 CPU cores, setting 99% frees one core. BOINC rounds down to the nearest fraction.

I wouldn't use that setting. Instead, when setting the <count> fields to control how many GPU tasks run at once I'd set the <avg_ncpus> fields such that when all GPUs had work there would be a CPU core freed, but if GPU work was running out all CPU cores would go back to doing pure CPU tasks. For a single GPU situation, <avg_ncpus> would be the same as or a tiny bit higher than <count>, for multiple GPUs it would scale down:

            1 GPU       2 GPUs       3 GPUs
<count>  <avg_ncpus>  <avg_ncpus>  <avg_ncpus>
  0.5        0.5          0.25         0.167
  0.33       0.34         0.167        0.112
  0.25       0.25         0.125        0.084
  0.2        0.2          0.1          0.067
                                                                   Joe

OK. I'm about to install a spanking new Radeon HD6570. Everything has been trucking along since the last update, with benchmarks around 6600 dhrystone. I still have the same CPU, and only have an interest in running SETI at the moment.

If I'm getting this right, I can run lunaticswin6440.04.exe, set it for CPU (NO GPU), and it should optimize for the new card. Am I missing anything?

You mean your Radeon HD6570? Then you should fill in the ATI part of the installer. Or do you want to only run CPU? (Check CPU capabillities like
SS(S)E3(x)4.1/4.2).
Why don't you use your 'spanking new Radeon HD6570', you can run MB
and AstroPulse work which works amazingly effective on an ATI 5000 or higher
series. A runtime of less then 1 hour to a few hours, CPU time is very
depending on % of RADAR-Blanking.

You can run multiple WUs per GPU and crashes are very rare,
AV scans(!) disable them for BOINC also WINDOWs updates are the most common reasons. The rev.390 MB app. and rev.555 app. for AstroPulse are stable and
haven't seen much errors from these. have a look at this
host
an i7-2600 and 2 ATI 5870 GPUs.


I freed up one 'thread', 0.5 core of an i7 and used BOINCs setting:
use 87.50 of CPU, I used your method in my app_info.xml trying
to improve loading of the 2nd ATI 5870 GPU.
But stopped running multiple WUs on a GPU, less effective, easier to compaire
run-times v.s NVidia GPUs. Although GPU load running 1 a GPU are 67% average.
(GPU 0, always has a higher load as GPU 1)!

---

How about I just leave things as they are and see if they work better on the new card?

If you "leave things as they are" (= wrong apps) you will generate errors.

Either return to stock/standard apps (Remove/Add project)
or run again the Lunatics Installer (Lunatics_Win64_v0.40_setup.exe)
and choose the proper apps - AMD for CPU and nothing for GPU




P.S.
ap_6.01r557_SSE2_331_AVX.exe work OK and is faster on my AMD Athlon(tm) 64 Processor 3500+
http://setiathome.berkeley.edu/forum_thread.php?id=67407&nowrap=true#1213064

OK. I'm about to install a spanking new Radeon HD6570. Everything has been trucking along since the last update, with benchmarks around 6600 dhrystone. I still have the same CPU, and only have an interest in running SETI at the moment.

If I'm getting this right, I can run lunaticswin6440.04.exe, set it for CPU (NO GPU), and it should optimize for the new card. Am I missing anything?

---

How about I just leave things as they are and see if they work better on the new card?

If you "leave things as they are" (= wrong apps) you will generate errors.

Either return to stock/standard apps (Remove/Add project)
or run again the Lunatics Installer (Lunatics_Win64_v0.40_setup.exe)
and choose the proper apps - AMD for CPU and nothing for GPU




P.S.
ap_6.01r557_SSE2_331_AVX.exe work OK and is faster on my AMD Athlon(tm) 64 Processor 3500+
http://setiathome.berkeley.edu/forum_thread.php?id=67407&nowrap=true#1213064

---

 


- ALF - "Find out what you don't do well ..... then don't do it!" :)
 

For CPU apps
You should choose the app labled AMD.

For GPU
Minimum Card with OpenCL cabability is HD 4350. (AMD OpenCL requirements)
This has nothing to do with Lunatics or the optimized apps.

Agreed, He most likely has an onboard GPU that is only capable of using the Brooks Hybrid app for AP WU's. running MB won't work at all

This is 100% the case. It seems like the problem is that I'm using a stone age graphics card. I'm planing to upgrade in the near future anyway, I can move the upgrade up by a few months. How about I just leave things as they are and see if they work better on the new card?

---

...
c) But the driver version may not matter (i.e. you can stay on 1.4.1720 if it works)
because for that older GPU the only app in the Lunatics Installer that may work is:
AP6_win_x86_SSE_Brook_r555

c) this looks promisting. I looked at kwsn.net for ap6_win?x86_ss_r55e with no hits. I googled it as well, nothing. Where could I find it? Is there an ancient history archive somewhere?
[/url]

The Lunatics v0.40 installer includes the AP6_win_x86_SSE_Brook_r555.exe file, you just need to choose it.

                                                                   Joe