Jump to content

Armbian for Amlogic S912

chocho

By chocho
in General Chat

 Share

Recommended Posts

gnthibault

gnthibault

Posted
Posted

@Shimon Awesome !

Thank you for the dtb file, I should receive my H96 pro+ from china by the end of this week or the next (I hope), I will also be able to run some tests for those who are interested.

Shimon

Shimon

Posted
Posted (edited)

Ok, the borrowed Alfa is almost certainly a DDR3 box and it looks like it has some single-threaded and memory performance issues (hopefully due to the dtb file used):

 

http://openbenchmarking.org/result/1704102-RI-1609084HA38

 

and a most suspicious memory performance result:

http://openbenchmarking.org/result/1704107-RI-1702269RI77

 

It also gets as hot as 80 deg. Celsius during 8-core operation. More results will be added retroactively to this post.

 

http://openbenchmarking.org/result/1704101-RI-1609236LO94

http://openbenchmarking.org/result/1704103-RI-1703186RI71

 

 

 

 

Edited by Shimon
finished adding other links
1
gnthibault

gnthibault

Posted
Posted

@Shimon Thank you for those tests, they look interesting. On the RAM test, the difference between streams benchmark and RAM benchmark however looks weird. I know the former one (streams), and find the result encouraging.

However, the benchmark for computations is clearly disappointing. The 2DFFT algorithm in particular should peform well with multiple cores, maybe your CPU is throttling, do you have a solution for monitoring that possible issue ?

I am now considering buying a set of radiator and a fan. Do you know if I can simply plugin a radiator like this one, originally designed for raspberry Pi, on the H96 pro+ board ?

 

Thank you for your work.

buvaluy

buvaluy

Posted
Posted

It is look like my beelink gt1(last debian) works as slow as bananapi_1(A20 fedora mainline)  on single core tasks...

They both use oracle jdk. Jsp execution gets same time and some other tasks too.

 

1 core A20 = 1 core S912 ?????

 

Does anybody know good java benchmark?

Shimon

Shimon

Posted
Posted
12 hours ago, gnthibault said:

maybe your CPU is throttling, do you have a solution for monitoring that possible issue ?

I am now considering buying a set of radiator and a fan. Do you know if I can simply plugin a radiator like this one, originally designed for raspberry Pi, on the H96 pro+ board ?

 

There could be some throttling at temps close to 80 but the main problem seems to be single-core performance. I didn't expect any difference between those boxes there, and yet, it looks like that H96 could be underclocked. For example 8-core performance scaled as if the frequencies never reached 1.5Ghz. 

 

 A repeat on a GT1 would be interesting; even my MX3 is palpably heavier (same cooling solution) and runs much cooler at idle.

 

As for the fan, no idea, have a look at some board photos in the freaktab threads.

 

9 hours ago, buvaluy said:

Does anybody know good java benchmark?

 

After installing phoronix-test-suite and php, you can simply do:

$  ./phoronix-test-suite benchmark java

1
Shimon

Shimon

Posted
Posted
1 hour ago, buvaluy said:

 

I've run ramspeed test

Temp during the test was like 59 *C

 

That particular benchmark is fixed at 2 threads by default. Try c-ray, primesieve, n-queens or smallpt to hit much higher temps. 

buvaluy

buvaluy

Posted
Posted

C-ray give 114.94 points result ( http://openbenchmarking.org/result/1704125-SANE-CRAYEMP79)

I expected A53 S912 core would be 2 times faster then A7 allwinner A20 core.(according to geekbench)

Shimon, what do you think about low memory test result.

I have Beelink GT1(SN:A912** -> AP6255)
Any idea which dbt I shoult try?

 

 

gnthibault

gnthibault

Posted
Posted
On 12/3/2016 at 10:37 PM, lvmc said:

Yes. I downloaded respective img from balbes150 link, wrote to sdcard.

 

To boot you have to insert sdcard, unplug power, hold the key near the capacitor / power plug, plug power and release key after some seconds.

 

On screen image will take a little bit to turn on.

 

@balbes150, nand-sata-install is writing upto 100%, but after reboot android is booting again, not the probably written armbian image.

 

Do you know what else we need to do?

 

Received my H96 pro+ one hour ago, Antutu benchmark looks pretty good, as reported by others (around 42K, I'll update this post with a picture if I can).

However, my box does not boot on the SDcard, and there does not seem to be a power switch on the box, what can I do ?

20170412_160137.jpg

20170412_181422.jpg

20170412_184218.jpg

gnthibault

gnthibault

Posted
Posted
18 minutes ago, buvaluy said:

You need update android to boot from sdcard, as I understand.

Try to use power on remote.

Android says no newer version detected (2017-02-27)

My remote controller has no battery unfortunately...

I tried to boot from both usb as well... nothing happened, may there is a problem with my image, how can I check that ?

buvaluy

buvaluy

Posted
Posted

I did it just one time.

I think you have to update android firmware - choose something like aml_autoscript in boot directory on sdcard. 

Try to find more info on first 2 pages

gnthibault

gnthibault

Posted
Posted

Ok, sorry for my silly question, it appeared that everything was actually working fine, I just forgot to use the toothpick method...

This is actually great, I put the ubuntu server image along with Shimon dtb, and everything seems to work fine. I'll focus on installing software during the next days/week, to finally setup my mini cluster of 4 H96 pro+ ^^

Shimon

Shimon

Posted
Posted
6 hours ago, buvaluy said:

 

C-ray give 114.94 points result ( http://openbenchmarking.org/result/1704125-SANE-CRAYEMP79)

I expected A53 S912 core would be 2 times faster then A7 allwinner A20 core.(according to geekbench)

Shimon, what do you think about low memory test result.

I have Beelink GT1(SN:A912** -> AP6255)
Any idea which dbt I shoult try?

Your c-ray result is completely expected, considering you didn't use CFLAGS=-Ofast.

As for your expectations, A53 is roughly 60% faster than Cortex A5 running 32-bit code and we're talking about a more powerful A7.

 

My guess is the dtb file I linked to (the only one to work with H96 Pro+ at the time) is somehow setting the cpu/memory frequencies too low or otherwise, something is seriously broken. 

 

Single core performance seems to correspond to about 1230 Mhz which happens to lie halfway between big cores' 1.5Ghz and little cores' 1Ghz.

@balbes150 I can't help noticing the names of the cpufreq governors start with arm-big-little... 

 

2 hours ago, gnthibault said:

I just forgot to use the toothpick method...

Or to use the Update&Backup app.

 

2 hours ago, gnthibault said:

to finally setup my mini cluster of 4 H96 pro+ ^^

Can you recommend a single PSU solution? 

gnthibault

gnthibault

Posted
Posted
2 hours ago, Shimon said:

Or to use the Update&Backup app.

[...]

Can you recommend a single PSU solution? 

 

I didn't know about this Update and Backup application, could be interesting to make a backup image from my last box this weekend (I insalled xenial on the emmc memory of the three other).

Sorry for the PSU, I thought about that for some time, but finally decided to remain in an ultra low cost /cheap setting, with 4 individual AC/DC adapter, plus the one for the Gigabit switch.

 

I may take some picture of the setup next week if you are interested anyway.

 

I will now try to run some homemade benchmark (some are still to develop) for single node GEMM and distributed GEMM, with NEON instructions enabled.

buvaluy

buvaluy

Posted
Posted

Somebody please run pts/ramspeed test. (at least 1)copy 1)integer)

I can't sleep calmly, I think something wrong with my beelink gt1 ....)

Shimon

Shimon

Posted
Posted

@buvaluy Nah, I'm sure it's fine. I don't have an S912 box any more so I can't try this idea myself but you could rerun a few benchmarks after issuing this command:

  

for i in `seq 4 7` ; do echo 0 | sudo tee /sys/devices/system/cpu/cpu$i/online ; done

shutting down the little cores.  (hopefully I got the numbering scheme right)

balbes150

balbes150

Posted
Posted

I have not studied this question, so this is just my assumption. Software or kernel detects 8 cores and try them all in parallel (the same) to use. I am inclined to think that the reason in the software, because other programs\testing (that you asked for at the time), show the full performance by using a basic kernel with a full load.

buvaluy

buvaluy

Posted
Posted

I've run tomcat site test. xxx/yyy - (java logic time) /(jsp-time[single thread]) 

 

8 cores

1543/1255 518/932 530/570 579/579 690/365 535/339 688/454 715/384
4 cores

915/530 755/570 572/448 715/285 457/401 453/357 508/261 394/259 467/354 398/191 326/146
 

Look like 4 cores really better especially jsp(single core) computing. 

 

Here is bananapi1(A20) results

965/489 856/558 865/326 1108/556 1044/319 851/532 821/301 851/588 735/284

Shimon

Shimon

Posted
Posted
5 hours ago, balbes150 said:

I am inclined to think that the reason in the software, because other programs\testing (that you asked for at the time), show the full performance by using a basic kernel with a full load.

 

Actually, I remember noticing a few pages back in c-ray, 8-core performance was not scaling as expected, so the issue was already there as well. We were trying to test HMP at the time, that's why nobody thought about doing single-threaded tests.

 

Turning off little cores makes the frequency scaling behave like any 4-core S905 system, including single-core performance. That's the problem (bug?) here, that is, as long as all cores are online, the frequency gets lowered, even if just 1-2 cores are being used. 

 

@buvaluy Could you try the most obvious test, activating just one little core to see what happens? Provided you already have just 4 cores active, add one more: 

echo 1 | sudo tee /sys/devices/system/cpu/cpu4/online

 

1
logol

logol

Posted
Posted

Hi everybody,

first of all i would like to thank you balbes150 for his great work.

I've managed to set up Armbian on my S912 TV-Box and run nginx/nextcloud on it. Nginx is also used as a reverse-proxy-server for my windows backend. Now, I try to implement a script which automatically send a wake on lan magic paket to the windows backend if needed. For this reason I would like to log the iptables. Unfortunately the kernel has not enabled the LOG module. To enable the LOG module the kernel has to be confired with CONFIG_IP_NF_TARGET_LOG enabled.
Is it possible to change the kernel? Please indulge me, i am pretty new in linux...

Thank you in advance for any help you can provide.
 

Shimon

Shimon

Posted
Posted
On 4/15/2017 at 0:25 PM, logol said:

To enable the LOG module the kernel has to be confired with CONFIG_IP_NF_TARGET_LOG enabled.

As well as these options, needed for running iotop: 

 - I/O accounting support (CONFIG_TASKSTATS, CONFIG_TASK_DELAY_ACCT, CONFIG_TASK_IO_ACCOUNTING)

 

On 4/15/2017 at 0:15 PM, buvaluy said:

cray 8core/5core/4core            114/166/195        (less is better)

ramspeed 8core/5core/4core  1162/1208/5000  (more is better)

Yeah, that 2nd benchmark utilises just 2 cores, so it's clear the number of active cores doesn't matter as long as just one little core remains online.  Amlogic should be asked to clarify what's going on here.  I sincerely hope we're not witnessing another S905 2GHz situation.

 

@buvaluy Could you try using the hotplug governor? It should be able to work around this limitation, giving us either 1-4 fast cores or 5-8 slower cores automatically. 

buvaluy

buvaluy

Posted
Posted

cray:

8core_hotplug     114

8core_ondemand 114

4core_ondemand 114

4core_hotplug     198

 

ramspeed

8core_hotplug     1460

8core_ondemand 1141

4core_ondemand 5133

4core_hotplug     5103

 

cpufreq-set -g hotplug, just set 4 first cores to hotpug mode, other stays ondemand.
 

Shimon

Shimon

Posted
Posted
15 minutes ago, buvaluy said:

cray:

4core_ondemand 114

That's a typo, 8 cores were definitely used. Other than that, it seems the unholy alliance of big.little scheduling and DVFS settings are causing this issue. 

 

You could either experiment with some cpufreq tunables or try rebuilding the kernel with:

`CONFIG_ARM_DT_BL_CPUFREQ=y`. Not sure if it's possible to disable `CONFIG_ARM_BIG_LITTLE_CPUFREQ` altogether.

 

https://github.com/150balbes/lib/blob/master/config/kernel/linux-amlogics905x-default.config#L451-L452

troyanz

troyanz

Posted
Posted

Hi all.
I try this build for TV BOX X98pro s912 3gb/32gb 
I boot from USB flash. All work fine, network, wifi, bluetooth
But firefox in youtube not play sound  I install chromium, work great.
Thanks for your work. I will be test this build some time.

x98-pro-amlogic-s912-3gb-32gb-android-6-0.jpg

Guest
This topic is now closed to further replies.
 Share
Go to topic listing
×
×
  • Create New...

Important Information

Terms of Use - Privacy Policy - Guidelines

  I accept