Pine 64 benchmarks

[g_t_j]

http://www.aliexpress.com/item/FREE-Ship...50220.html
http://www.aliexpress.com/item/2-pcs-lot...87879.html

this is reasonable alternative if you can wait few days for shipping.
If you can wait a bit longer I would wait for heatsink that @tllim is working on. A large heatsink covering the SoC, memory and PMIC (with or without a larger fan) would be the best solution.

Good one. Thanks for the info.

I got some of those copper SDRAM heatsinks and this is the reason I asked if they could be used on the Pine.
I will probably wait and get the official heatsink which, from what I understand, will be similar to Odroid's one, covering more than just the processor's chip.

[tllim]

           

http://www.aliexpress.com/item/FREE-Ship...50220.html
http://www.aliexpress.com/item/2-pcs-lot...87879.html

this is reasonable alternative if you can wait few days for shipping.
If you can wait a bit longer I would wait for heatsink that @tllim is working on. A large heatsink covering the SoC, memory and PMIC (with or without a larger fan) would be the best solution.

Here are the heat sink samples photos, Aluminum with copper plated. We will need to characterize the heat sink and will takes few week.

[janjwerner]

Are these heatsinks http://www.ebay.co.uk/itm/191167775314?_...EBIDX%3AIT suitable for the SDRAM chips? I have a few of those available.

I'm also considering attaching something like this http://www.ebay.co.uk/itm/172088180004?_...EBIDX%3AIT on the processor's chip. I've read that the A53 chip itself is 14mm x 14mm square so one of those (15 x 15) above should do the job.

The pure solid copper plate is the good heat sink.

Yep, I think so. This is why I restricted my market search to pure copper heatsinks.

The question is how am I going to attach the above plate on the processor's chip? Do I have to use a drop of superglue?

please no super glue!
use a thermal pad: http://www.amazon.com/s/ref=nb_sb_noss_2...hermal+pad

[norm24]

Ebay 221836873756 and 221830353784 is with thermopad

[Andrew2]

Here are the heat sink samples photos, Aluminum with copper plated. We will need to characterize the heat sink and will takes few week.

Looks good and efficient but I wonder whether you take care of the slightly different height of DRAM and SoC?

[Andrew2]

BTW: Anyone interested in this topic (ARMv8 code with all optimisations turned on, sane dvfs settings, relationship between heat dissipation and performance, relationship between clockspeed, heat and stability and so on... everything except of benchmarks!   ) might have a look here: http://linux-sunxi.org/User:Tkaiser#Reli..._on_Pine64

Back to benchmarks: In the following graph we used cpuminer (rather efficient and using NEON optimised code) to get an idea how throttling affects performance: With a fan close to the SoC's heatsink and clockspeed adjusted to 1200MHz we get stable 3.440 khash/s. After switching the fan off the SoC's temperature increases and performance decreases since maximum clockspeeds will be lower to prevent overheating (these are pretty conservative settings, you could also increase the thermal tresholds but it makes not that much sense):



What this graph does not show is how important software/settings for benchmark scores are since three weeks ago this benchmark would've shown stable ~830 khash/s instead. Why? Since back then longsleep's OS images relied on Allwinner's default throttling settings (killing CPU cores and never bringing them back). One week ago the results without a fan would've also been drastically lower since we just recently improved throttling behaviour again by using more intermediate steps which automatically improves peak performance under constant full load. And this is still WiP, see the link below.

What's also interesting when using 'benchmarks that matter' [TM] (with all compiler/code optimisations turned on and not random stuff that runs partially only single threaded as common on openbenchmarking.org) is how better A64 performs when compared to eg. H3 (used on Orange Pis for example): https://github.com/longsleep/build-pine6...-196267397 (same clockspeed, no throttling -- without monitoring this stuff in parallel you just fool yourself when testing since you get random results that are more influenced by things like ambient temperature than anything else)

[Andrew2]

And a few final words for those thinking about number crunching, clustering or any other high performance stuff and putting their device into an enclosure. Do you remember the 68°C from before when running the cpuminer benchmark with a small fan blowing air over the cheap heatsink? It get's even better:



64/65°C in the beginning dropping down to 62/63°C after using two more stripes of adhesive tape to prevent unwanted ventilation to the sides. I used now a larger fan and control the airflow much more. This 'enclosure' is not just beautiful but also highly efficient when it's about improving heat dissipation:



I hope you get the idea. When you use a large fan mounted on the bottom of an enclosure containing 5 boards and you do enclosure design well then you take care about controlling the airflow so that as much air as possible 'hits' the heatsink's surfaces.

While I doubt that enclosure makers will ever get the idea to let convection jump in (mount everything vertically -- use large heatsinks) and that airflow is the most important factor for today's SBCs/SoCs if they should perform constantly under high load at least I will build one the next time for this sort of tests. A 120mm fan mounted vibration free is almost silent and more efficient than annoying smaller ones. And performance improves since throttling starts later and when no throttling is involved for reasons I still don't understand absolutely performance also increases with lesser temperatures:

3440 khashes/sec when running 5°C hotter compared to an average 3490:

[g_t_j]

Are these heatsinks http://www.ebay.co.uk/itm/191167775314?_...EBIDX%3AIT suitable for the SDRAM chips? I have a few of those available.

I'm also considering attaching something like this http://www.ebay.co.uk/itm/172088180004?_...EBIDX%3AIT on the processor's chip. I've read that the A53 chip itself is 14mm x 14mm square so one of those (15 x 15) above should do the job.

The pure solid copper plate is the good heat sink.

Yep, I think so. This is why I restricted my market search to pure copper heatsinks.

The question is how am I going to attach the above plate on the processor's chip? Do I have to use a drop of superglue?

please no super glue!
use a thermal pad: http://www.amazon.com/s/ref=nb_sb_noss_2...hermal+pad

I won't. I will be buying some thermal pad instread.

[BarBrian]

Hello !

Tllim:
When board dimensions and all have been settled, could you contact Flirc for a enclosure ? https://flirc.tv/more/raspberry-pi-case
The magic here is that the molded aluminium case itself works as a heatsink, having an internal aluminium "tentacle" to reach
the ASIC with thermal pad.

https://www.google.com/search?q=flirc+ca...cMvAO3M%3A

BarBrian

[stealthbanana]

Rather than starting a new thread, though I would just add to this.

When I get a new comp, I like to run a basic benchmark, using the simple primes test from sysbench, keeping the command line the same.

sysbench --test=cpu --cpu-max-prime=20000 run

(as per the the example in the man page.)

Must admit I am either well impressed, or dubious about the Pine64 result, could someone else recheck it for me?

As can be seen by the results below, the Pine is over 40x faster at computing primes than an original Raspberry Pi B and even faster than my laptop.  Not in anyway scientific, but all tests ran using Debian, or a Debian derived distro.

Raspberry Pi B (single core)
total time taken by event execution:  1321.6702

Raspberry Pi 2 (quad core)
total time taken by event execution: 769.8208

Banana Pi (Dual Core)
total time taken by event execution:  734.1427

Orange Pi Plus (Quad Core)
total time taken by event execution:  477.5902

Orange Pi PC (Quad core)
total time taken by event execution:  456.3728

And the Pine64 (quad core 2GB)
total time taken by event execution: 31.7711       (ran several times to verify the result)




(And as a test, on a few x86 systems

Netbook - (Intel Atom N270)
total time taken by event execution: 186.9946

Desktop (AMD FX-8350 - 8 core)
 total time taken by event execution: 14.0435

Laptop - (Intel i3 Dual core with Hyperthreading)
total time taken by event execution: 39.7917