DPM on Radeon HD6570

DPM on Radeon HD6570

[Boszormenyi Zoltan]

Hi,

I read this Phoronix article:
http://www.phoronix.com/scan.php?page=article&item=amd_hd6000_dpm&num=1

Congrats to the progress achieved so far.

However, I can see an interesting deviation for HD6570 from the
observed trend of other chips.

r600g can reach 80+ percent of the performance of Catalyst
for most HD6xxx chips except for 6570, where the performance
is around 10-20 percent.

Do you have a theory about this difference?
Maybe DPM doesn't work as intended on HD6570?

I have this kind of video card, so I wanted to test it myself.
The exact model of my card is:
http://www.sapphiretech.com/presentation/product/?cid=1&gid=3&sgid=1087&pid=1176&psn=&lid=1&leg=0#

I have installed kernel 3.11-rc6 on Fedora 19 using this koji kernel:
http://koji.fedoraproject.org/koji/buildinfo?buildID=457463

I haven't tested Catalyst but my r600g results mostly match
the ones in the article even with a different CPU.

Best regards,
Zoltán Böszörményi

Re: DPM on Radeon HD6570

[Alex Deucher]

On Wed, Aug 21, 2013 at 10:31 AM, Boszormenyi Zoltan <zboszor@pr.hu> wrote:
> Hi,
>
> I read this Phoronix article:
> http://www.phoronix.com/scan.php?page=article&item=amd_hd6000_dpm&num=1
>
> Congrats to the progress achieved so far.
>
> However, I can see an interesting deviation for HD6570 from the
> observed trend of other chips.
>
> r600g can reach 80+ percent of the performance of Catalyst
> for most HD6xxx chips except for 6570, where the performance
> is around 10-20 percent.
>
> Do you have a theory about this difference?
> Maybe DPM doesn't work as intended on HD6570?
>

Are you seeing the same results on your board?  If so are the results
roughly the same with dpm enabled vs. disabled?  If so I doubt there
is a problem with dpm.  On older dGPUs like this one dpm won't really
improve performance since the cards come up with relatively high
clocks by default.  It's mainly for saving power when the GPU is idle.

> I have this kind of video card, so I wanted to test it myself.
> The exact model of my card is:
> http://www.sapphiretech.com/presentation/product/?cid=1&gid=3&sgid=1087&pid=1176&psn=&lid=1&leg=0#
>

Note that a lot of 6570 cards, including yours use DD3 memory rather
than GDDR5 so they will have fairly limited memory bandwidth.

Alex

> I have installed kernel 3.11-rc6 on Fedora 19 using this koji kernel:
> http://koji.fedoraproject.org/koji/buildinfo?buildID=457463
>
> I haven't tested Catalyst but my r600g results mostly match
> the ones in the article even with a different CPU.
>
> Best regards,
> Zoltán Böszörményi
>
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: DPM on Radeon HD6570

[Grigori Goronzy]

On 21.08.2013 16:31, Boszormenyi Zoltan wrote:
> Hi,
>
> I read this Phoronix article:
> http://www.phoronix.com/scan.php?page=article&item=amd_hd6000_dpm&num=1
>
> Congrats to the progress achieved so far.
>
> However, I can see an interesting deviation for HD6570 from the
> observed trend of other chips.
>
> r600g can reach 80+ percent of the performance of Catalyst
> for most HD6xxx chips except for 6570, where the performance
> is around 10-20 percent.
>
> Do you have a theory about this difference?
> Maybe DPM doesn't work as intended on HD6570?
>

There are some ways to check if DPM functions correctly.

The kernel log (dmesg) contains a dump of the PowerPlay tables, search 
for "power state". It's possible that tables aren't read correctly or 
flaky to start with.

You can also monitor the precise runtime power level (clocks and 
voltages) of the GPU with /sys/kernel/debug/dri/0/radeon_pm_info. Maybe 
dynamic switching does not work for some reason, and the GPU always uses 
the lowest level even under load. Just run some demanding OpenGL app and 
check that file to see if the GPU switches the power level.

You should probably file a bug on bugs.freedesktop.org (product DRI, 
component DRM/radeon) and attach dmesg output and anything else that 
might be useful.

Best regards
Grigori

Re: DPM on Radeon HD6570

[Boszormenyi Zoltan]

Hi,

thanks for your response.

2013-08-21 17:39 keltezéssel, Alex Deucher írta:
> On Wed, Aug 21, 2013 at 10:31 AM, Boszormenyi Zoltan <zboszor@pr.hu> wrote:
>> Hi,
>>
>> I read this Phoronix article:
>> http://www.phoronix.com/scan.php?page=article&item=amd_hd6000_dpm&num=1
>>
>> Congrats to the progress achieved so far.
>>
>> However, I can see an interesting deviation for HD6570 from the
>> observed trend of other chips.
>>
>> r600g can reach 80+ percent of the performance of Catalyst
>> for most HD6xxx chips except for 6570, where the performance
>> is around 10-20 percent.
>>
>> Do you have a theory about this difference?
>> Maybe DPM doesn't work as intended on HD6570?
>>
> Are you seeing the same results on your board?  If so are the results
> roughly the same with dpm enabled vs. disabled?  If so I doubt there
> is a problem with dpm.  On older dGPUs like this one dpm won't really
> improve performance since the cards come up with relatively high
> clocks by default.  It's mainly for saving power when the GPU is idle.

I have enabled dpm:
$ cat /proc/cmdline
BOOT_IMAGE=/vmlinuz-3.11.0-0.rc6.git0.1.fc20.x86_64 
root=UUID=00df37a2-be3d-46fe-963a-ca08977fc5f6 ro quiet rhgb radeon.audio=1 radeon.dpm=1 
LANG=hu_HU.UTF-8

I have just tried "openarena 0.8.5" again with forced "low" performance.
Results is 26.87fps with low performance, 59.40-59.70fps with forced high
performance.

>
>> I have this kind of video card, so I wanted to test it myself.
>> The exact model of my card is:
>> http://www.sapphiretech.com/presentation/product/?cid=1&gid=3&sgid=1087&pid=1176&psn=&lid=1&leg=0#
>>
> Note that a lot of 6570 cards, including yours use DD3 memory rather
> than GDDR5 so they will have fairly limited memory bandwidth.

I know. The 6570 tested by Phoronix must be GDDR5
but it's not mentioned specifically.

Best regards,
Zoltán Böszörményi

>
> Alex
>
>> I have installed kernel 3.11-rc6 on Fedora 19 using this koji kernel:
>> http://koji.fedoraproject.org/koji/buildinfo?buildID=457463
>>
>> I haven't tested Catalyst but my r600g results mostly match
>> the ones in the article even with a different CPU.
>>
>> Best regards,
>> Zoltán Böszörményi
>>
>> _______________________________________________
>> dri-devel mailing list
>> dri-devel@lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: DPM on Radeon HD6570

[Alex Deucher]

On Wed, Aug 21, 2013 at 1:49 PM, Boszormenyi Zoltan <zboszor@pr.hu> wrote:
> Hi,
>
> thanks for your response.
>
> 2013-08-21 17:39 keltezéssel, Alex Deucher írta:
>
>> On Wed, Aug 21, 2013 at 10:31 AM, Boszormenyi Zoltan <zboszor@pr.hu>
>> wrote:
>>>
>>> Hi,
>>>
>>> I read this Phoronix article:
>>> http://www.phoronix.com/scan.php?page=article&item=amd_hd6000_dpm&num=1
>>>
>>> Congrats to the progress achieved so far.
>>>
>>> However, I can see an interesting deviation for HD6570 from the
>>> observed trend of other chips.
>>>
>>> r600g can reach 80+ percent of the performance of Catalyst
>>> for most HD6xxx chips except for 6570, where the performance
>>> is around 10-20 percent.
>>>
>>> Do you have a theory about this difference?
>>> Maybe DPM doesn't work as intended on HD6570?
>>>
>> Are you seeing the same results on your board?  If so are the results
>> roughly the same with dpm enabled vs. disabled?  If so I doubt there
>> is a problem with dpm.  On older dGPUs like this one dpm won't really
>> improve performance since the cards come up with relatively high
>> clocks by default.  It's mainly for saving power when the GPU is idle.
>
>
> I have enabled dpm:
> $ cat /proc/cmdline
> BOOT_IMAGE=/vmlinuz-3.11.0-0.rc6.git0.1.fc20.x86_64
> root=UUID=00df37a2-be3d-46fe-963a-ca08977fc5f6 ro quiet rhgb radeon.audio=1
> radeon.dpm=1 LANG=hu_HU.UTF-8
>
> I have just tried "openarena 0.8.5" again with forced "low" performance.
> Results is 26.87fps with low performance, 59.40-59.70fps with forced high
> performance.
>

Sounds like you are refresh rate limited.  Try disabling
swapbufferswait in your xorg.conf:

Section "Device"
        Identifier  "card0"
        Driver      "iradeon"
        Option "SwapbuffersWait" "false"
EndSection

and disable vsync in the 3D driver, set env var:
vblank_mode=0

Also what does the performance look like with dpm disabled?

Anyway, it doesn't sound like dpm is an issue.

Alex

>
>>
>>> I have this kind of video card, so I wanted to test it myself.
>>> The exact model of my card is:
>>>
>>> http://www.sapphiretech.com/presentation/product/?cid=1&gid=3&sgid=1087&pid=1176&psn=&lid=1&leg=0#
>>>
>> Note that a lot of 6570 cards, including yours use DD3 memory rather
>> than GDDR5 so they will have fairly limited memory bandwidth.
>
>
> I know. The 6570 tested by Phoronix must be GDDR5
> but it's not mentioned specifically.
>
> Best regards,
> Zoltán Böszörményi
>
>
>>
>> Alex
>>
>>> I have installed kernel 3.11-rc6 on Fedora 19 using this koji kernel:
>>> http://koji.fedoraproject.org/koji/buildinfo?buildID=457463
>>>
>>> I haven't tested Catalyst but my r600g results mostly match
>>> the ones in the article even with a different CPU.
>>>
>>> Best regards,
>>> Zoltán Böszörményi
>>>
>>> _______________________________________________
>>> dri-devel mailing list
>>> dri-devel@lists.freedesktop.org
>>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>
>

Re: DPM on Radeon HD6570

[Boszormenyi Zoltan]

2013-08-21 18:30 keltezéssel, Grigori Goronzy írta:
> On 21.08.2013 16:31, Boszormenyi Zoltan wrote:
>> Hi,
>>
>> I read this Phoronix article:
>> http://www.phoronix.com/scan.php?page=article&item=amd_hd6000_dpm&num=1
>>
>> Congrats to the progress achieved so far.
>>
>> However, I can see an interesting deviation for HD6570 from the
>> observed trend of other chips.
>>
>> r600g can reach 80+ percent of the performance of Catalyst
>> for most HD6xxx chips except for 6570, where the performance
>> is around 10-20 percent.
>>
>> Do you have a theory about this difference?
>> Maybe DPM doesn't work as intended on HD6570?
>>
>
> There are some ways to check if DPM functions correctly.
>
> The kernel log (dmesg) contains a dump of the PowerPlay tables, search for "power 
> state". It's possible that tables aren't read correctly or flaky to start with.

Here it is:

[    2.252567] [drm] Internal thermal controller with fan control
[    2.252619] == power state 0 ==
[    2.252621]  ui class: none
[    2.252622]  internal class: boot
[    2.252624]  caps:
[    2.252625]  uvd    vclk: 0 dclk: 0
[    2.252627]          power level 0    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.252628]          power level 1    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.252629]          power level 2    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.252630]  status: c r b
[    2.252632] == power state 1 ==
[    2.252632]  ui class: performance
[    2.252633]  internal class: none
[    2.252635]  caps:
[    2.252636]  uvd    vclk: 0 dclk: 0
[    2.252637]          power level 0    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.252640]          power level 1    sclk: 40000 mclk: 80000 vddc: 1000 vddci: 0
[    2.252641]          power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
[    2.252642]  status:
[    2.252643] == power state 2 ==
[    2.252644]  ui class: none
[    2.252645]  internal class: uvd
[    2.252646]  caps: video
[    2.252647]  uvd    vclk: 70000 dclk: 56000
[    2.252651]          power level 0    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
[    2.252652]          power level 1    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
[    2.252653]          power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
[    2.252654]  status:
[    2.257099] switching from power state:
[    2.257105]  ui class: none
[    2.257107]  internal class: boot
[    2.257109]  caps:
[    2.257111]  uvd    vclk: 0 dclk: 0
[    2.257113]          power level 0    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.257114]          power level 1    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.257116]          power level 2    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.257117]  status: c b
[    2.257129] switching to power state:
[    2.257130]  ui class: performance
[    2.257132]  internal class: none
[    2.257133]  caps:
[    2.257135]  uvd    vclk: 0 dclk: 0
[    2.257137]          power level 0    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
[    2.257139]          power level 1    sclk: 40000 mclk: 80000 vddc: 1000 vddci: 0
[    2.257140]          power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
[    2.257141]  status: r
[    2.258555] [drm] radeon: dpm initialized


>
> You can also monitor the precise runtime power level (clocks and voltages) of the GPU 
> with /sys/kernel/debug/dri/0/radeon_pm_info. Maybe dynamic switching does not work for 
> some reason, and the GPU always uses the lowest level even under load. Just run some 
> demanding OpenGL app and check that file to see if the GPU switches the power level.

Just running this loop:

[root@localhost ~]# while `/bin/true` ; do cat /sys/kernel/debug/dri/0/radeon_pm_info ; 
sleep 1 ; done

while the terminal is sufficiently large on the screen results in changing power levels:

uvd    vclk: 0 dclk: 0
power level 0    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 0    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 0    sclk: 10000 mclk: 15000 vddc: 900 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
uvd    vclk: 0 dclk: 0
power level 2    sclk: 65000 mclk: 80000 vddc: 1050 vddci: 0
^C

>
> You should probably file a bug on bugs.freedesktop.org (product DRI, component 
> DRM/radeon) and attach dmesg output and anything else that might be useful.
>
> Best regards
> Grigori
>

So, it seems DPM works well but the card is weak.
Sorry for the noise.

Best regards,
Zoltán Böszörményi

Re: DPM on Radeon HD6570

[Boszormenyi Zoltan]

2013-08-21 19:55 keltezéssel, Alex Deucher írta:
> On Wed, Aug 21, 2013 at 1:49 PM, Boszormenyi Zoltan<zboszor@pr.hu>  wrote:
>> Hi,
>>
>> thanks for your response.
>>
>> 2013-08-21 17:39 keltezéssel, Alex Deucher írta:
>>
>>> On Wed, Aug 21, 2013 at 10:31 AM, Boszormenyi Zoltan<zboszor@pr.hu>
>>> wrote:
>>>> Hi,
>>>>
>>>> I read this Phoronix article:
>>>> http://www.phoronix.com/scan.php?page=article&item=amd_hd6000_dpm&num=1
>>>>
>>>> Congrats to the progress achieved so far.
>>>>
>>>> However, I can see an interesting deviation for HD6570 from the
>>>> observed trend of other chips.
>>>>
>>>> r600g can reach 80+ percent of the performance of Catalyst
>>>> for most HD6xxx chips except for 6570, where the performance
>>>> is around 10-20 percent.
>>>>
>>>> Do you have a theory about this difference?
>>>> Maybe DPM doesn't work as intended on HD6570?
>>>>
>>> Are you seeing the same results on your board?  If so are the results
>>> roughly the same with dpm enabled vs. disabled?  If so I doubt there
>>> is a problem with dpm.  On older dGPUs like this one dpm won't really
>>> improve performance since the cards come up with relatively high
>>> clocks by default.  It's mainly for saving power when the GPU is idle.
>> I have enabled dpm:
>> $ cat /proc/cmdline
>> BOOT_IMAGE=/vmlinuz-3.11.0-0.rc6.git0.1.fc20.x86_64
>> root=UUID=00df37a2-be3d-46fe-963a-ca08977fc5f6 ro quiet rhgb radeon.audio=1
>> radeon.dpm=1 LANG=hu_HU.UTF-8
>>
>> I have just tried "openarena 0.8.5" again with forced "low" performance.
>> Results is 26.87fps with low performance, 59.40-59.70fps with forced high
>> performance.
>>
> Sounds like you are refresh rate limited.  Try disabling
> swapbufferswait in your xorg.conf:
>
> Section "Device"
>          Identifier  "card0"
>          Driver      "iradeon"
>          Option "SwapbuffersWait" "false"
> EndSection
>
> and disable vsync in the 3D driver, set env var:
> vblank_mode=0

vblank_mode was already 0 before:

[zozo@localhost ~]$ cat .drirc
<driconf>
     <device screen="0" driver="r600">
         <application name="Default">
             <option name="fthrottle_mode" value="2" />
             <option name="pp_celshade" value="0" />
             <option name="pp_jimenezmlaa" value="0" />
             <option name="always_have_depth_buffer" value="false" />
             <option name="pp_jimenezmlaa_color" value="0" />
             <option name="pp_nogreen" value="0" />
             <option name="force_glsl_extensions_warn" value="false" />
             <option name="pp_nored" value="0" />
             <option name="disable_glsl_line_continuations" value="false" />
             <option name="vblank_mode" value="0" />
             <option name="allow_large_textures" value="1" />
             <option name="pp_noblue" value="0" />
         </application>
     </device>
</driconf>

I have added this:

[root@localhost xorg.conf.d]# pwd
/etc/X11/xorg.conf.d
[root@localhost xorg.conf.d]# cat 99-vblank.conf
Section "Device"
     Identifier    "card0"
     Driver        "radeon"
     Option        "SwapbuffersWait" "false"
EndSection

With forced high performance, I got 118.73 fps. Wow. :-)

> Also what does the performance look like with dpm disabled?

Same as forced high with dpm enabled.

> Anyway, it doesn't sound like dpm is an issue.

Indeed, dpm works as intended.
Something was misconfigured at Phoronix then.

Thanks for resolving this for me,
Zoltán Böszörményi

>
> Alex
>
>>>> I have this kind of video card, so I wanted to test it myself.
>>>> The exact model of my card is:
>>>>
>>>> http://www.sapphiretech.com/presentation/product/?cid=1&gid=3&sgid=1087&pid=1176&psn=&lid=1&leg=0#
>>>>
>>> Note that a lot of 6570 cards, including yours use DD3 memory rather
>>> than GDDR5 so they will have fairly limited memory bandwidth.
>> I know. The 6570 tested by Phoronix must be GDDR5
>> but it's not mentioned specifically.
>>
>> Best regards,
>> Zoltán Böszörményi
>>
>>
>>> Alex
>>>
>>>> I have installed kernel 3.11-rc6 on Fedora 19 using this koji kernel:
>>>> http://koji.fedoraproject.org/koji/buildinfo?buildID=457463
>>>>
>>>> I haven't tested Catalyst but my r600g results mostly match
>>>> the ones in the article even with a different CPU.
>>>>
>>>> Best regards,
>>>> Zoltán Böszörményi
>>>>
>>>> _______________________________________________
>>>> dri-devel mailing list
>>>> dri-devel@lists.freedesktop.org
>>>> http://lists.freedesktop.org/mailman/listinfo/dri-devel