I’ve talked about CPU cooling and how important it is before. Here’s a couple of videos from “ExplainingComputers.com” showing the effects of different heat sinks on a R.Pi3 and R.Pi4.
His benchmark program, sysbench, is OK for this use, but realize that on a 64 bit OS it shifts over to use of the GPU and is much more efficient. Were I doing this, I’d make my own CPU loading program that didn’t do that so I could get consistent results regardless of the OS used.
This matters as in another of his videos he was confused as to why the Odroid XU4 fared poorly on sysbench when compared to another board on a sysbench benchmark, completely missing that the 32 bit XU4 was running in the CPUs and the other (Rock64 IIRC) was using the GPUs. But for this use it doesn’t really matter as he’s just making the chip hot with it.
This is also, though, why you need to compare the same OS across boards if at all possible. Some OSs know how to use the GPU, others not so much… IIRC, Raspbian is still v7 instructions so runs in the CPU while Debian is v8 and uses the GPUs for Sysbench.
So why no R.Pi M4 for me? Simples. It has more heat production that the M3 and does a lousy job of heat removal. The M3 is marginal on heat removal, but it’s OK with a modest heat sink for “average” uses. For heavy computes, one would need to put on a really big heat sink, or go to active cooling. That adds about $15 to $20 to the cost. At $55 total, I can get an 8 core Odroid XU4 with the heat sink already installed from Ameridroid. So tell me again why I want to go through the trouble and fuss and expense of a DIY on the Pi when I can get a really nice solution already done at the same cost?
Yes, it is 32 bit. But most real world problems are fine in 32 bit. On actual CPU benchmarks, it performs very nicely. Now if you happen to have a GPU suited problem, using the GPUs will matter. But very few people do, or know how to make them “go”. So the GPU 64 bit Sysbench benchmark is really just blowing smoke up their skirts as they will never get that themselves.
If I ever really needed that, I can get the RockPro64 with more memory, better memory / IO design, and a faster chip, also with a designed for purpose heat sink (though bought separately) for only a little more money. In a prior posting I did in fact do that.
Now I’ve compared the RockPro64 to the Odroid XU4 and for all the stuff I tend to do they are “about the same” with the small exception of a few single CPU intensive things (like some web pages). On a day to day basis, I like the XU4 the most (partly due to the software choices available – Devuan) and it simply does everything I really need. Rarely does it load up all 8 cores, and rarely is is really bogged down on a single big core at 100% on a single thread task. Basically, it’s “enough computer”. And it is silent. No fan. (I got the Q model).
With that in mind, look at the work involved to get adequate cooling for the Pi M3 without a fan, or for the Pi M4. Should I ever get a R. Pi M4, I might try the shim fan and see if it really is quiet enough, or pay up the $20 for the Ice Tower and not connect the fan. That is a huge heat pipe system. Almost as big as the Pi itself.
Oh, and remember that (at least as of a while back when last I checked) Raspbian was running v7 32 bit mode AND you had to specifically reset the CPU clock to get it to run full speed. What’s the point of 64 bit hardware if you are only using it in 32 bit mode? What’s the point of a 1.4 GHz top, if you configure the default to limit at 900 MHz? One hopes they have fixed that now. (I’ve been using Devuan the last couple of years so not had those issues. I’m running real 64 bit full clock on my Pi M3, with heat sink).
But the fact that I HAD to deal with the question of CPU limit setting and what instruction set (v7 vs v8) was the default OS using was “sand in the teeth”. Most folks buying a Pi and just putting Raspbian on it, are really getting about 1/2 or 1/3 of the performance they thought they were buying. Then, if you use it, it rapidly heat limits even at that! Don’t get me wrong, I love my PiM3. But that is AFTER applying an after market heat sink, changing the CPU limiter settings, and installing a real 64 bit OS (Devuan).
Compare the XU4: Open box. Apply power. Use.
OK, enough of my tech talk, here’s the videos. Pi M3 cooling first:
At about 81 C the Pi heat limits. Then it goes slower to avoid overheating. Note that the graph showed a small heat sink with fan stays under about 62 C; so that’s a fine performance. His extreme passive heat sink stays under 50 C, so is also fine performance. But look what it took to get there! Sawing slabs of copper? Building custom enclosures…
Here’s the Pi M4 video. In it he first runs it in the default plastic case. The one with zero ventilation. ( I have one. With a Pi M3 with small heat sink I must leave the top off for even modest use loads.) He then works his way up through several heat sink types.
I have 2 complaints about his comments. First, notice that in the chart at 1:30 the times of two choices complete in “about 11 minutes” for a best case 10 minute run. That’s only a very small amount of heat limited performance. For almost any normal use cases, that just won’t happen. It may show up in heavy compute loads, but even there very few real programs use 100% of ALL cores for 10+ minutes. So for my purposes, I’d rank all of them as “good enough for all practical purposes”. Also note that the default board in the default case is 41 minutes to complete. It is heat limited to about 1/4 of expected performance…
The other complaint is when he “talks dirt” about using the Ice Tower with the fan off and praises the value of “moving some air”. Well, yes, you get lower numbers; BUT, as a passive cooler it is well below the point where you heat limit or hurt anything and it IS silent. That is stellar performance! Not something to talk down.