Passively-cooled CPU Thermals (Part 2)

In the first Passively-cooled CPU Thermals post I looked at how the completely fanless Streacom DB4 I built performed under various CPU loads.  A question over on Silent PC Review asked how the thermal curve looked when the CPU was unloaded.  Good question — let’s find out!

It seems logical that a passively-cooled computer that uses heat pipes and large slabs of aluminium will heat up and cool down differently than an air-cooled system with a (relatively) small fin stack.

When I performed the previous round of tests, I ended up with graphs like this:

Thermals 60m at 100%

Now, one thing that niggled away at the back of my mind was how the temperature seemed to spike at the start and then flatten off relatively quickly.  Given that I was sampling the temperature sensors at 5s intervals, I wasn’t sure if the sudden change in angle was just due to a lack of resolution, or because of something else.  Hmmm…

Anyway, the SPCR question gave me an excuse to run another test but instead of putting the CPU under load and heating the system up, this time I’d be unloading the CPU and cooling the system down.  I wondered if/how that would be different.

Test 5 — CPU unloaded from 100% at an ambient temperature of 20°C

The Ryzen 5 1600 was first placed under a 100% load (all 12 threads pegged at 100%) for about an hour until an equilibrium temperature was reached (60⁰C).  The load was then removed and the CPU temperature recorded:

Thermals 60m cooldownMy previous set of results used wider graphs, but when websites scale them
the text ends up a bit small and blurry, so this time I made the graph narrower.
Otherwise the testing and recording setup was identical.

The CPU cooled down from 60⁰C to 49⁰C almost instantly, and then gradually made its way down to 34⁰C over the course of an hour.  It would have cooled down a couple of degrees more, but I wasn’t prepared to wait — that sudden change in angle was even more pronounced and needed investigating.

Obviously 1 hour graphs sampled every 5s were too coarse to shed a whole lot of light on what was happening in those first few seconds, so I had to run some more granular tests.  I figured 1 minute sampled every 1s would do the trick.

Test 6 — CPU loaded to 100% for 60s at an ambient temperature of 20⁰C

First 60s heating up

Test 7 — CPU unloaded from 100% for 60s at an ambient temperature of 20⁰C

First 60s cooling down

Nope, it wasn’t a figment of my imagination.  The sudden change in angle is definitely there — both when heating up and cooling down.

Let’s put the results ‘side-by-side’:

First 60s up and down

What we’re seeing are two different response curves.  When the CPU is initially (un)loaded the temperature changes at ±1⁰C every 1s for about 8–9s.  It then flattens out dramatically and changes by ±1⁰C every 16–24s after that.  Huge difference.

Changes in thermal conductivity can usually be explained by changes in the material or media being used.  In this case we have a CPU, soldered to a copper IHS, plated in nickel, covered by a thin layer of TIM, covered by a copper shim, covered by TIM, clamped against an aluminium block with exposed copper heat pipes.

This image (from Streacom’s DB4 Manual) might make the stacking a bit clearer:

CPU Cooling Stack

If we ignore the thin layers of TIM, it’s pretty-much metal all the way through to the heat pipes before we reach the first medium that could possibly have a 16–24x lower thermal conductivity — the water inside the heat pipes themselves.

I suspect that the thermal mass between the heat pipes and the CPU can buffer about 8–9s worth of heat output, and that — due to high thermal conductivity — this mass heats up and cools down rather quickly.  The water in the heat pipes, however, needs to undergo a phase change (from liquid to vapour), travel to the end of the heat pipe, undergo another phase change (from vapour to liquid), and then flow back — a process that could easily be an order of magnitude slower at transferring heat.

Interesting stuff.

At the end of the day it doesn’t matter exactly how thermally conductive your cooling system is — as long as it’s conductive enough to get rid of all of the heat your CPU is producing without impacting performance (i.e. thermally throttling), it’s fine.

The previous tests (and now weeks of daily hammering) show that the DB4+LH6 combination is easily capable of cooling a stock Ryzen 5 1600 — no matter how heavily you load it or how long your jobs run for.  You can run a Ryzen 5 1600 at 100% load all day, every day, and it won’t even break a sweat.

AMmmD and Streacommm — I’m lovin’ it!  😉

19 thoughts on “Passively-cooled CPU Thermals (Part 2)

  1. Hiya! Thanks for the write-up, it’s fantastic to see that passive cooling is finally practical (without having to build a crazy case from scratch).

    That’s a very neat observation on the CPU cooling curves. I think the curve behaviour is more about thermal masses than conductivity. The temperature sensor you’re reading is on the CPU die, which is physically small so has a small thermal mass. Then the heatsink, the heatpipes and the case all have larger masses of their own. When you’re dumping lots of energy into the die, it gets hotter than the heatsink, and energy flows down the temperature gradient into the ‘sink. Then when the heatsink is above ambient that energy flows down the next gradient to the case.
    As soon as you turn it off the die rapidly equilibriates with the much-massier heatsink. Then the second part of the curve is probably the heatsink itself equilibriating with the case, which is happening at the same time. From your graphs can guess that at 100% load there’s 8-9C difference from the CPU die to the heatsink and pipes, and the remainder is the total heatsink-ambient difference. One thing this is telling you is that your CPU->heatsink thermal resistance is hovering around 0.1°C/W which is pretty healthy!

  2. You hit it on the head: Practical passive cooling without needing crazy cases (or modding). I don’t mind building stuff, but small case construction requires tools and skills I just don’t have, and demands tolerances that my eyes and big, clumsy hands just can’t deliver. The DB4 comes in a box with crystal-clear instructions (well done Streacom!), can easily be assembled in just a few hours, and performs exactly as advertised. What more can you ask for, really?

    As this is my first purely passive build, I’m not used to seeing thermal curves like this. Learning things as I go. Can you explain what you mean by “One thing this is telling you is that your CPU->heatsink thermal resistance is hovering around 0.1°C/W which is pretty healthy!” I’m happy to hear that it’s healthy, but I’m not sure where the 0.1°C/W number comes from.

  3. Yeah, it’s *almost* tempting enough for me to replace my workstation ^_^

    I was a bit sloppy with the maths in that comment. You’re seeing jumps of, by my reading, perhaps 8-9C in the first slope? Most of that is the die/heatsink temperature difference settling in or out. Based on a TDP of 65W for your CPU that eyeballs like ~8.5/65 = 0.14°C/W. Of course that assumes you’re actually pumping 65W in there, which probably means getting all the vector units going and anything else as well, so that figure might be a bit optimistic.

    Handwaving maths aside, I think the best piece of information from your data is that such a machine should do okay (with the LH6 kit, at least) even on those delightful summer days where it tops out at 44 indoors…

  4. Ah, got it. Thanks. The 0.14°C/W figure makes more sense now.

    FYI: The power draw at the wall is ~35W when idling and ~75W when all 12 threads are on 100% load. My loads are mainly fluid simulations — so aren’t likely to stress the CPU to the same levels as a purely synthetic torture test would. I’m happy with that — I prefer real world tests.

    I use air-conditioning to keep the ambient temperature well below 30°C when I’m using the computer. Can’t imagine working at one at 44°C — although I can see that sort of temperature being quite possible if a DB4 was running as a server in a confined space.

  5. Thanks for the interest Honker. I’m currently in the process of building a house, so that’s been consuming most of my spare time lately. Once windows and bushfire shutters are sorted out I should have more time on my plate to continue tests. I haven’t forgotten.

  6. Hi, do you have any data when the cpu has a TDP of 95W or even 105W?

  7. No (or at least, not yet).

    I plan on overclocking the Ryzen 5 1600 when I get a chance — just to see what happens to the thermals — and (based on other reports) that should push the power consumption to ~95W. No promises on when that will happen, though — I’ve currently got a fair bit on my plate.

    I do not have a 105W CPU to put into this system, nor do I plan to ever get one.

  8. Hi, this is somewhat unrelated to this specific post, but do you think it’s possible to mount the LH6 Kit on the GPU to get the GPU to spread to 2 sides to accomodate higher TDP cards?

  9. The LH6 kit contains two spreaders and three long heat pipes. It does not contain a cooler or copper shim that you can clamp onto the GPU. The GPU kit does, but the cooler only has four slots and thus can only handle four heat pipes. Those four heat pipes are already bonded to a single wall. A single wall is capable of dissipating the heat output from four heat pipes before diminishing returns makes connecting any more of dubious value. In other words if you get a GPU kit then it’s already maxxed out — whether you bond it to a single wall or two walls. Even if you used a LH6 (or something else) to spread the heat making it’s way from the four-slot GPU kit over two walls, it wouldn’t make much of a difference — because you still only have four heat pipes and that’s the bottleneck.

    To support a higher TDP GPU you need a larger cooler with a shim and more slots for heat pipes — like the 6-slot stock one that comes with the DB4 to cool the CPU, or something from a third-party.

    Then there’s the length of the LH6 heat pipes to consider. The only wall that isn’t bonded to something is the back wall. The LH6 heat pipes — even though they are longer — are not long enough to run all the way from the GPU, back past the motherboard, to the back wall, with enough length to have meaningful contact with the spreader.

    So, for the above two reasons, you cannot make use of a LH6 to enhance the cooling of the GPU kit to support a GPU with a higher TDP. Nor can you just use a LH6 by itself. You would need to purchase some quite long third-party heat pipes and custom bend them yourself. You would also need at least a 6-slot cooler and shim.

    That said, it has been done before. I think a guy on the SFF Forums managed to get a GTX 1070 in there — but it wasn’t easy and the whole system runs really hot. If you are prepared to put in the effort, and don’t mind gambling with the lifespans of your components, then it can be done. I suspect such a card would thermally throttle on a regular basis, though — which kind of defeats the purpose of installing a powerful card in the first place.

  10. Hello!
    Just wanted to drop you a line and say how grateful I am for your study. It is the best source I could find with actual data on a real world heat pipe design. The info you provided has been invaluable to me as I am designing an enclosure for a $25k PCBA. Thanks again!

  11. Rick, I’m glad you found the data useful! This was the first totally passive system I’ve built and used, so thermal testing was a bit more ‘involved’ than it was in the past.

    The system has now been in daily use for four and a half years, and has clocked up over 24,000 hours of use in that time. A fair chunk of that was with the GPU at 100% and the CPU at about 40%. Everything still works perfectly fine. No coil while (or any other noises) have developed — it’s still totally silent. Apart from a minor and very early “shutdown on idle” issue that was quickly resolved with a BIOS (AGESA) update, it’s been rock solid. If the cooling solution was deficient in any way, I think I would know by now.

    $25k is a huge amount of money, but I can think of no better way to protect such a sum than with a fundamentally simple and passive design that eliminates all moving parts from the equation. Solid state all the way!

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s