Jump to content

orangepione H3 an experiment with small heatsink

ag123

By ag123
in Allwinner sunxi

 Share

Recommended Posts

ag123

ag123

Posted
Posted

orange pi one H3 idles somewhat warm in part due to a power regulator that switches only between 1.1 and 1.3 v. orange pi pc could idle at pretty low temperatures (closer to room without a heat sink due to a better power regulator as well), and the h3 runs rather hot say > 70deg C when under loads

the trouble is the SOC die is 14x14mm, and if one wants to use a large heat sink, it wouldn't fit and you would need to use the copper shim method, to provide some clearance to fit a large heat sink

 

that is troublesome and hence , i made do and made an attempt to try out *small* heat sinks, e.g.

https://www.aliexpress.com/item/10pcs-Computer-Cooler-Radiator-Aluminum-Heatsink-Heat-sink-for-Electronic-Chip-Heat-dissipation-Cooling-Pads-14/32890197245.html

the dealer provided some thermal tapes, it is convenient, but i did not use them, instead i used some ordinary heat sink compounds like this

https://www.ebay.com/sch/i.html?_from=R40&_nkw=hy510&_sacat=0&_sop=15

the results looks like this, i used a little too much thermal compound and they ooze from the edges

pioneheatsink1.thumb.jpg.691d1d5e3fdd4e5bba6c8f258ae9b579.jpg

 

tests

platform: debian stretch mainline kernel (switched to development / nightly builds) Linux orangepione 4.14.68-sunxi #161 SMP

 

to evaluate how well such small heat sink works i made it (H3) do some math (square matrix multiplication) based on codes adapted from

https://github.com/mtrebi/matrix-multiplication-threading.git

this probably won't give the best mflops, gflops but it is able to run all 4 threads concurrently doing the matrix multiplication

the test does a multi-threaded 1000x1000 matrix multiplication concurrently, this is 2N^3 flops (2 billion floating point ops)

 

here are the results

orange pi one

 

no heat sink idle

  Reveal hidden contents

heat sink idle

  Reveal hidden contents

 

i noted that the idle temperatures are about 5 degrees lower average vs without the heat sink

 

no heat sink load 5 cycles of 1000x1000 matrix multiplication (each cycle takes about 10-11 secs hence about 1 min run)

  Reveal hidden contents

 

heat sink load 5 cycles of 1000x1000 matrix multiplication (each cycle takes about 10-11 secs hence about 1 min run)

  Reveal hidden contents

 

ok so that is 180mflops (no heat sink) vs 214mflops (heat sink), note that in this particular case, i started off at lower temperatures (i.e. after the board just started), hence, there is more thermal headroom which results in a higher difference. this is quite reflective of a transient load.

 

no heat sink load 10 cycles of 1000x1000 matrix multiplication (each cycle takes about 10-11 secs hence about 2 minutes run)

  Reveal hidden contents

 

heat sink load 10 cycles of 1000x1000 matrix multiplication (each cycle takes about 10-11 secs hence about 2 minutes run)

  Reveal hidden contents

 

ok so that is 178 Mflops no heat sink vs 194 Mflops, with heat sink

no heat sink

960000    6162    6162    6162    6162 <<< cpu spend most of the time here
1008000    3307    3307    3307    3307 <<< cpu spend most of the time here

vs heat sink

1008000    6528    6528    6528    6528 <<< cpu spend most of the time here
1200000    7056    7056    7056    7056  <<< cpu spend most of the time here

this case would more likely reflect sustained loads

 

conclusion

the small heat sink is probably inadequate but nevertheless it made a modest improvement

the cpu is able to spend more of the time at the higher frequencies under load vs without heat sink, hence the Mflops improvement

for transient conditions (i.e. short high loads) this small heat sink apparently made a visible difference as seen in the Mflops improvement

somewhat lower idle temperatures (for orange pi one), orange pi pc could do without a heat sink at idle

ag123

ag123

Posted
  • Author
Posted

out of curiosity i fine tuned the source codes making use of the quad execute for floating point on cortex a7 (simply unroll the loop)

https://community.arm.com/processors/f/discussions/5277/cortex-a7-pipeline-is-non-symmetric-what-does-this-attribute-mean

  Reveal hidden contents

2389.49Mflops !

 

now that looks like the power of a modern arm cortex a7 processor. my guess is the gains are from pipelining efficiencies and the fp quad execute

 

the command to build is c++ -O2 -std=gnu++11 -pthread -o mat main.cpp

then run ./mat

sources attached

main.cppFetching info...

 

yes that is the orange pi one or orange pi pc (H3) mainline kernel

and for all that speeds the temperatures (with heatsink) looked uneventful (no high temps etc)

soctemp1.png.3e933bf4be1575f5a7b0e130286298cb.png

 

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