Cooling The Raspberry Pi 4 With The Fan SHIM & FLIRC For Better Performance

[danmcgrew]

Not to stir things back up again, after this finally wound down, but do you realize that RK3399 is only dual-A72 + quad-A53, while the N2's S922X is quad-A73 + dual-A53?

I'm just asking, because I had forgotten the RK3399 only had two A72 cores.

And according to this, the A73 is about 35% faster, per clock (i.e. DMIPS/MHz), at least for integer workloads.

https://en.wikipedia.org/wiki/Compar..._ARMv8-A_cores


If you knew this, fine. If not, now you do. I'm not trying to change your mind - just wanted share this info for anyone who's interested.

Yep; knew that.

And I'm STILL going to buy a Pinebook Pro 14.1", 1080p IPS display Linux laptop--with full privacy controls AND a magnesium-alloy case; NOT PLASIC--for $199.95.


https://www.cnx-software.com/2019/07...-linux-laptop/

[coder]

In the case of RK3399,

Not to stir things back up again, after this finally wound down, but do you realize that RK3399 is only dual-A72 + quad-A53, while the N2's S922X is quad-A73 + dual-A53?

I'm just asking, because I had forgotten the RK3399 only had two A72 cores.

And according to this, the A73 is about 35% faster, per clock (i.e. DMIPS/MHz), at least for integer workloads.

https://en.wikipedia.org/wiki/Compar..._ARMv8-A_cores


If you knew this, fine. If not, now you do. I'm not trying to change your mind - just wanted share this info for anyone who's interested.

[mvdvarrier]

From the sound of it, your solution works. I just mentioned that for future reference, should you have occasion to run across other boards with heatsinks that can be scavenged.

Ok Thank you Sir.

[coder]

It is Brass. I have no other option.

From the sound of it, your solution works. I just mentioned that for future reference, should you have occasion to run across other boards with heatsinks that can be scavenged.

[tuxd3v]

I generally prefer passive cooling, when I can use it, because it's quieter, less prone to dust buildup, and mechanical parts are often the first points to fail.

For example, I sometimes use a machine that just has the standard Intel cooler that came with the CPU. I rarely load more than one of its cores, so this is normally fine. However, I took it out of sleep, one day, and the fan just didn't spin up. I got all kinds of warnings that the CPU was throttling and then I found that the fan wasn't spinning. No explanation. Rebooted and it came back just fine. That happened about 2 years ago and never had any issues with it since. Clearly, fans (PWM, even more so) just add another point of failure. If it were passively-cooled, that couldn't have happened.

I understand,
And the fan could even fail with time( and it will fail for sure with time.. )..
If that happens, the passive cooling will have to enter, and start lowering the voltage/Clocks, to adjust temp..

But I understand, that a complete passive solution, if working, is ideal..

In the case of RK3399, there are people that have them running, [email protected],[email protected],
In a complete passive solution but they have a massive heating, and the board is vertically mounted, so that the heated air could go up, and pass in the heatsink more cooler air, coming from bellow..

What does this have to do with anything? None of the hardware we're discussing does that, to my knowledge. It sounds like you're talking about Intel's Turbo Boost.

If you consider that cheating, then you can just look at the base clocks of CPUs and I think most BIOS will even let you disable it. But that's completely off-topic, here.

That behaviour were reported in previous AMLogic Socs( S905, and others.. ), they worked at bigger frequencies during some time, but because they could not sustain the heat, after some time they were forced to lower the clocks..

But again, I am not saying that this behaviour happen in s922x( it should be a improved Soc.. ),
It could even be, that the cooling solutions that the previous Socs used, were simply badly designed, and the problem was not with the SoCs themselves..

Well, we'd know if you actually look at the link from my post. They state the ambient temperature above and directly in the heading of the graph, and you can read the x-axis to see that the tests went for 35000 seconds (about 9.7 hours).

Here, I'll quote it for you:
Note that they ran both stress-ng and glmark2, at the same time!

If they run on that situations, my bad, I read the tests, but I doesn't toke note for the time they run it.
My Apologies for that..

It's not trapped. The design is somewhat open, allowing for convection. It also has large fins on the sides. And whether there's trapped air or not, the steady state is below the throttling point. So, their thermal solution seems to be entirely adequate.

It could be,
And the half node, will also help on that, but this boards( rockPi4 and N2 ), to take maximum efficiency on passive cooling, they should run in the vertical position..
Because in that way, you grant that, is less hot air trapped bellow( to maximise the air flow..since the hot air goes up, and it pushes cooler air bellow to pass in the heat sink, more often, to make heat transfer more effective.. ).

Any Way,
Any board discussed is a nice board!
Each has its pros and cons, even raspberry pi4 has its pros( and also its cons..ie: they limited the Soc to 1.5Ghz, because they knew that running it at 1.8 or more would be outside limits for the size of the board, and cooling solution.., also they messed up in the USB type C conector.. ).

regards,

[coder]

Wen I said that RockPi4 in my opinion was superior to N2,

Okay, that's great. Let's move on.

I just want to clarify a couple points.

Active cooling should always take precedence..

I generally prefer passive cooling, when I can use it, because it's quieter, less prone to dust buildup, and mechanical parts are often the first points to fail.

For example, I sometimes use a machine that just has the standard Intel cooler that came with the CPU. I rarely load more than one of its cores, so this is normally fine. However, I took it out of sleep, one day, and the fan just didn't spin up. I got all kinds of warnings that the CPU was throttling and then I found that the fan wasn't spinning. No explanation. Rebooted and it came back just fine. That happened about 2 years ago and never had any issues with it since. Clearly, fans (PWM, even more so) just add another point of failure. If it were passively-cooled, that couldn't have happened.

As an example, you cannot advertise that you have a core that works at lets say 2.5Ghz, but it can only do that during 2 seconds( an example.. )
For me that is not a core that works at 2.5Ghz.. I call that cheating

What does this have to do with anything? None of the hardware we're discussing does that, to my knowledge. It sounds like you're talking about Intel's Turbo Boost.

If you consider that cheating, then you can just look at the base clocks of CPUs and I think most BIOS will even let you disable it. But that's completely off-topic, here.

We don't know for example during how much time the tests were done, what was the room temperature, and a lot of other parameters like, orientation of the SBC, and so on..

Well, we'd know if you actually look at the link from my post. They state the ambient temperature above and directly in the heading of the graph, and you can read the x-axis to see that the tests went for 35000 seconds (about 9.7 hours).

Here, I'll quote it for you:

Thermal characteristics
To check the thermal throttling, we ran some heavy CPU and GPU loads together on the SoC and monitored temperature. We ran the test within a chamber that keeps the ambient temperature at 35°C. So we believe the large stock passive heatsink is enough for very heavy computing loads.

cmd: stress-ng –cpu 6 –cpu-method matrixprod && glmark2-es2-fbdev –off-screen –run-forever

Note that they ran both stress-ng and glmark2, at the same time!

Since they don't say what orientation it had, I would assume the standard orientation. However, it's a fair point.

You need to remember that, the CPU transfer heat to the heatsink, but the heatsink needs to exchange that heat with the air around..
If it his bellow( which is the case.. ), a lot of hot air is trapped beneath, and so, the heatsink continues to heat, each time more gradually..

It's not trapped. The design is somewhat open, allowing for convection. It also has large fins on the sides. And whether there's trapped air or not, the steady state is below the throttling point. So, their thermal solution seems to be entirely adequate.

[tuxd3v]

But your comparison of RAM speed to core speed is unjustifiable.
Look at the ODROID benchmarks vs. their N1. Real benchmarks beat your facile analysis.

I Can't because you are comparing apples to oranges..
You can't produce 2 boards, then compare them, and after say that your second board is faster than another vendor board..

That is a false test, by another words..its playing low..

It's funny how you started out trying to compare performance and now you're saying it's not valid?
Of course it's valid, for someone who doesn't need the RockPro64's extra capabilities. However, for someone who does need them, then there's not much point in it, because it's a given which one they'll buy.

That was my point,
Wen comparing rpi4, rockpi4,N2..

Wen I said that RockPi4 in my opinion was superior to N2,
But I put there a observation,
That in power consumption,maybe N2 consumes less power due to half the node size...this feature was not even in your Odroid internal related tests, its one that should be..

I'm not sure you understand thermodynamics. Their temperature data clearly shows a plateau, which means the heatsink has reached steady state and is able to dissipate all the additional heat it's adsorbing from the SoC.

This is not cheating. This is what all modern SoCs do, in order to avoid burning out. That's why you also need to look at the clock speed graph that ODROID posted right below their temperature graph, which shows no throttling.

I understand that very well..

The thing is,
When the air around is heated, it will go up, the cold one is in the room floor, and the heated one, goes to the roof..
The problem is that, that board doesn't let the warm air go up, so in the heat transfer, the air around is already heated... is not efficient..

Like I said before try to run it full throttle during a day( but without cheating, on throttling the cores...full throttle, like I do on rockpro64 for 14 months nonstop, in a room at 36C..a72@2Ghz,a53 @1.5Ghz fixed frequency!! no more than 56 degrees ).

Not all vendors cheat on performance, but AmLogic is known to do that( at least in Android, for the benchmarks.. )..
That also doesn't mean that they continue to do that( I just don't trust much, due to their historic past.. )..

The trip points, they exist of course..
Are a passive cooling mechanism, nowadays all boards implement it,

But that trip points are above the normal temperatures( in the superior limit.. , or at least should be ), and the SoC should implement emergency shutdown if above determined temperature..
But this 2 things are a separate subject.. they are related with safety temperatures that you can operate your cores without passive cooling, and the trip point limits to do passive cooling..

Active cooling should always take precedence..

As an example, you cannot advertise that you have a core that works at lets say 2.5Ghz, but it can only do that during 2 seconds( an example.. )
For me that is not a core that works at 2.5Ghz.. I call that cheating

There are a lot of factors that affect performance, besides RAM speed. Not all tasks have the same amount of memory I/O, and cache varies between SoCs. That's why I say you need to just look at the benchmarks.

Again, this is you not understanding thermodynamics. Heat conduction is not dependent on orientation, and that's how heat is transferred into the N2's heatsink. But you don't even need to understand that much - just look at the data. Regardless of why or how, the data clearly shows the N2 doesn't throttle.

I agree, there are lot of factors, but if you have a fast cpu and slower memory, its a bottleneck..

The only thing, that tests, reflects are comparisons between N1 and N2, from Odroid..
We don't know for example during how much time the tests were done, what was the room temperature, and a lot of other parameters like, orientation of the SBC, and so on..

You need to remember that, the CPU transfer heat to the heatsink, but the heatsink needs to exchange that heat with the air around..
If it his bellow( which is the case.. ), a lot of hot air is trapped beneath, and so, the heatsink continues to heat, each time more gradually..

Do that experience yourself..
I am telling you this because I also have a RockPi4, with a *huge* heatsink( bellow offcourse because the CPU is bellow ), and I also notice that..
I need to use a fan on the side to push fresh air, or cooler, to the bottom of the board..
You only notice this if you do hard stress tests, otherwise, has a desktop, it should be fine..

[tuxd3v]

But the N2 is faster, being based on A73 cores, while the RockPi64 uses A72 cores. ODROID made a N1 that was based on the same SoC as the RockPi64, but they cancelled it.

Define Faster?

The N1 was cancelled because at the time ODROID canoot compete on same memory frequencies, and it experienced a shortage of them..
Logically using lower frequency memories( which they continue to use.. ), their board has less performance..

Comparing N2, with N1, and then making a parallel with Rockpro64, is NOT valid..
Also the tests I saw are finite in time..

Try to run it, in a room at 36-38C, at full throttle during 24 hours..

[mvdvarrier]

Good point.

Yes, exactly. Use a decent heat sink and decent thermal compound.

Perhaps you know this, but try to use the least amount of thermal compound possible, as long as it covers the entire intersection between the SoC and heat sink. Using more thermal compound than necessary will actually reduce thermal conductivity, since it's not as conductive as the metal of the heat sink.

BTW, perhaps the heat sink you're using is copper? Copper is the best (solid) heat sink material, while aluminum is second best and more common (since it's cheaper).

It is Brass. I have no other option.

[coder]

Let's not forget to appreciate that some level of power management is actually implemented, which is why it can work without a heatsink at all.

Yes, that's the reason for the performance discrepancies, shown throughout the article.

Decent drivers, hereunder thermal throttling, is a luxury you don't always get with even expensive SBCs. My WandPi managed to overheat when idling, despite the generous heatsink that came with it. I measured 107C (heatsink temperature) with IR thermometer.

IMO, such garbage belongs in the waste bin. A lot of these smaller-volume products don't have proper driver support. I think that's what you encountered.