(03-30-2016, 01:19 PM)peterz Wrote: [ -> ]I personally would rather see some big heatsink covering all 4 chips (with no fan if possible).
But somebody on this forum wrote height of all 4 chips is not equal.
Tllim is working with some vendor on making a heatsink.
Here is a enclosure for Raspberry PI made from aluminium with "cube" which connects enclosure body and surface of CPU. That might be the best passive heatsing from all the options (once are dimensions modified for PINE64).
http://raspi.tv/2014/the-flirc-case-for-raspberry-pi-b
Yeah they are definitely different height, and with the curved boards, that presents a challenge. Also the 2Gb versions have ram chips on the bottom of the pcb as well.
(03-30-2016, 01:36 PM)rahlquist Wrote: [ -> ] (03-30-2016, 01:19 PM)peterz Wrote: [ -> ]I personally would rather see some big heatsink covering all 4 chips (with no fan if possible).
But somebody on this forum wrote height of all 4 chips is not equal.
Tllim is working with some vendor on making a heatsink.
Here is a enclosure for Raspberry PI made from aluminium with "cube" which connects enclosure body and surface of CPU. That might be the best passive heatsing from all the options (once are dimensions modified for PINE64).
http://raspi.tv/2014/the-flirc-case-for-raspberry-pi-b
Yeah they are definitely different height, and with the curved boards, that presents a challenge. Also the 2Gb versions have ram chips on the bottom of the pcb as well.
This stuff might help with the uneven heights and curved boards:
http://www.amazon.com/viscous-thermal-pa...B00UTX7K2E
Unlike regular thermal pastes, which are designed to flow under higher heat to fill in very small spaces between a chip and heat sink that's in tight contact with the top of the packaging through capillary action, this is very thick (like silly putty), very sticky (may not need clamps, depending on the application), and doesn't change phase when heated. It's meant to replace thermal pads. Of course, it's not going to transmit heat as efficiently as direct contact with high-grade thermal compounds, but it is designed to fill in larger gaps between a chip and heat sink.
I just received some and plan to try it out when I've got some time.
(03-31-2016, 08:59 AM)patrickhwood Wrote: [ -> ] (03-30-2016, 01:36 PM)rahlquist Wrote: [ -> ] (03-30-2016, 01:19 PM)peterz Wrote: [ -> ]I personally would rather see some big heatsink covering all 4 chips (with no fan if possible).
But somebody on this forum wrote height of all 4 chips is not equal.
Tllim is working with some vendor on making a heatsink.
Here is a enclosure for Raspberry PI made from aluminium with "cube" which connects enclosure body and surface of CPU. That might be the best passive heatsing from all the options (once are dimensions modified for PINE64).
http://raspi.tv/2014/the-flirc-case-for-raspberry-pi-b
Yeah they are definitely different height, and with the curved boards, that presents a challenge. Also the 2Gb versions have ram chips on the bottom of the pcb as well.
This stuff might help with the uneven heights and curved boards: http://www.amazon.com/viscous-thermal-pa...B00UTX7K2E
Unlike regular thermal pastes, which are designed to flow under higher heat to fill in very small spaces between a chip and heat sink that's in tight contact with the top of the packaging through capillary action, this is very thick (like silly putty), very sticky (may not need clamps, depending on the application), and doesn't change phase when heated. It's meant to replace thermal pads. Of course, it's not going to transmit heat as efficiently as direct contact with high-grade thermal compounds, but it is designed to fill in larger gaps between a chip and heat sink.
I just received some and plan to try it out when I've got some time.
Never heard of the product before. Also I wouldnt want to use TIM material to make up those kind of gaps, proper machining maybe.
(03-30-2016, 10:36 AM)jhgoodwin Wrote: [ -> ]For the HDFS, I also was planning on getting some higher performance microsdxc modules, with specs like this or better:
http://www.amazon.com/PNY-Turbo-Performa...B00W77C1QU
Thoughts?
Forget about theoretical sequential transfer speeds since Pine64 won't exceed ~22MB/s anyway. Compare the
random I/O speed ratings of your card with
better ones (and keep in mind that cards of different size might perform different if they use a good controller -- the latter is the reason why I don't buy Kingston, PNY or any other vendor that does not produce own controllers and NAND dies, since you never know what you get, they exchange the controller between two batches and random I/O differs by 300% between batches).
And then please forget about the whole Hadoop idea or better check why it's important to use cluster nodes with ECC RAM for these sorts of storage clusters. If you still want to use cheap SBCs for this stuff seriously consider downclocking the DRAM and do extensive tests before (not possible with A64 ATM, when it's possible it will look like
this)
Regarding heatsinks numbers are available in Linux Development forum, you can combine even cheap ones with somewhat controlled airflow but the whole approach is pretty useless when we're talking about Hadoop anyway (I/O bound so the A53 cores will be more or less idle). But simply get 160 of
these and use 5 each for every Pine64 (SoC, 4 x DRAM, PMU doesn't need a heatsink in your 'mostly idle' scenario)
(03-30-2016, 01:19 PM)peterz Wrote: [ -> ]Here is a enclosure for Raspberry PI made from aluminium with "cube" which connects enclosure body and surface of CPU. That might be the best passive heatsing from all the options
This looks nice and is quite expensive so once you bought it you already
feel that it works perfectly (that's how psychology works). But if you start to measure then things look different. I've provided a script that sets up a
whole monitoring solution within minutes that can be also used to monitor the performance effects of throttling and results are simple (and the same as with any other chip on this planet)
If you really want to do number crunching you need some ventilation. And most importantly a somewhat working heatsink on the SoC (DRAM and PMU doesn't matter in this regard). The most expensive heatsink won't outperform the combination of a cheap one combined with some airflow. One large silent fan mounted below a bunch of Pine64 with controlled airflow is pretty enough to keep A64 below 80°C so no throttling happens even with heavy workloads (Hadoop is none)
I was thinking about getting separate thermal adhesive for all the heatsinks, but i'm not sure i want to commit to having them "permanently" on the board. It may be the only way to keep them on for a long time though.