Discussion Thermal Interface Materials for Consideration in Cooling Upgrades

Soul of Jacobeh

Administrator
Staff member
Administrator
Joined
Sep 10, 2018
Messages
27
Likes
18
#1
The first thing that comes to mind when you're trying to cool something more is a bigger, badder heat sink, or a liquid cooler, or what kind of radiators you're going to use. That's all fine and dandy, but you can often improve your thermal performance just as much by simply replacing your thermal interface. A thermal interface material is the component (usually paste) that goes between your device (either its naked lid, or its Integrated Heat Spreader - IHS) and your heat sink.

There's a few different kinds ranging from pastes to pads made out of metal, and they all have various properties that make them better suited for some scenarios over others, or make them cost-attractive.

In the list below, you'll see a number followed by "W/m". This is the unit of measurement for thermal conductivity. Basically, it's how well the material conducts heat from another object to itself, and can change based on how warm or cold it is. A cold material will typically better conduct heat than it would if it was already warm. This is why, sometimes, it's better to increase airflow than it is to get a larger heat sink. Returning the heat sink closer to ambient is often better than a larger heat sink that remains saturated.
I'm going to focus on interfaces designed for conducting at least 50W (like a CPU) over a 30x30mm area.
  • Thermal paste is anywhere from 2 - 12W/m in all directions.
    • +Good, general purpose, easy-to-apply, reliable application. Plenty of room for error.
    • +Can be further engineered for special application methods that allow for some exceptional tolerances or precision, as necessary.
    • ++Exceptionally cheap, even for the highest end of pastes (although retail pricing ruins this).
    • +Well-suited for precision-machined heat-sinks that mate very closely to the surface its cooling.
      • Sometimes it's better to use an extremely thin layer of paste than it is to use a better material that's thicker.
      • In this scenario, you're relying on direct metal-to-metal contact between a well-machined heat sink and a lid or IHS. The paste only serves to cover extremely small, infrequent gaps. If, instead, you separated the sink and device with something even as good as Indium, you might lose performance. Most metals conduct hundreds of W/m, given good contact.
    • -Lowest thermal conductivity of all common TIMs, although it's usually 'good enough'.
  • Graphite thermal interface pads conduct 35W/m between it and another surface. Across/within itself, it approaches 400W/m.
    • +The ~ 400W/m lateral heat spread means it handles hotspots very well - it rapidly distributes localized heat across itself to better displace it to the heatsink.
    • -Requires moderately high compression from the heat sink to conduct well.
    • -Electrically conductive.
    • -Not super easy to apply (it likes to slide around and is very light; you can't see if you've brushed it out of place while affixing the heat sink).
      • Because it is electrically conductive, you want to be extra careful and maybe double check if it applied well. It will often stick to the sink so be careful you don't lose it when removing the sink.
    • +Excellent, cost-effective upgrade for most consumer-grade heat sinks. I've found even a $20 Hyper 212 Evo acts like a $60 cooler if you replace the paste with a graphite pad.
  • Liquid metal is usually around 32 - 73W/m between surfaces. I do not know the lateral heat spread but I'd imagine it's excellent.
    • -Very difficult application.
    • -Electrically conductive - don't mess up.
    • -Depending on the metals it's interfacing, it may fuse them together or damage them.
    • -Good luck cleaning that stuff up.
    • -Sorta expensive.
    • +Precise application guarantees exceptional performance.
  • Indium Corporation "Heat Spring" thermal interface pads exceed 86W/m conductance, and - I think - exceeds the lateral heat transfer of graphite.
    • ++In basically all thermal regards, it exceeds the performance of graphite pads.
    • --Requires very high compression.
    • --Exceedingly cost prohibitive. Whereas a 40mm graphite pad is around $10 - $15 USD, an Indium pad is above $200.
      • Can be bought in bulk sheets to bring this cost down.
    • -Easy to damage.
    • -Precision application tools that require a bit of practice. (It's a glorified suction cup on a stick.)
  • Carbon NanoTube structures perform above Indium, at an estimated 3x that of Indium. Its real benefit is it performs excellently despite the usual contact resistance of the two materials it is interfacing.
    • +++Basically the realistic maximum we can achieve in thermal performance, given the more recent application and manufacturing processes.
      • Even something like the boiling action in two-phase immersion cooling can destroy the CNTs.
    • ---Untouchably expensive. In a situation where Indium might cost as little as $50 and as much as $200, CNTs would likely cost over $1200 or more.
    • --Brittle. Easily destroyed if mishandled, misapplied, or otherwise harmed by the given application
    • ---Unreasonably difficult to apply. Requires the manufacturer to apply it to your substrate directly, in most cases.
There's a few more thermal interfaces for specialty scenarios, such as ones that are highly heat conductive while remaining dielectric (electrically non-conductive), but they're beyond impractical for most scenarios. If you're in a situation that calls for these, you probably already know how to hunt them down and apply them. The major problem with most of these is that they're either highly reactive (will destroy or bond with your device or sink) or are extremely hard (and thus resistant to the compression that is necessary to effectively mate them to your device and sink).

I bring all this up both because it's interesting and because we're talking about FPGAs here that can push 300A through a core that's under 40x40mm. That's a lot of power (and a lot of heat) in a very small area. It'll come to a point where many cooling solutions are being limited by even high-grade paste. I believe one manufacturer I'm working with has mentioned a silver-heavy paste (not like Arctic Silver, although it does use small nano particles of silver) that I'm hoping to get some data on. There will likely be a "this is the most cost effective TIM for everybody" consensus in the future between the enthusiasts in the community. I imagine if it's one of the more expensive or difficult-to-use products, another group-buy could be arranged. (Personally, I'd be ecstatic if everyone and their mother was using Indium.)

--
Latest edit to correct errors in stated thermal performance of liquid metal and Indium.
 
Last edited:
OP
OP
Soul of Jacobeh

Soul of Jacobeh

Administrator
Staff member
Administrator
Joined
Sep 10, 2018
Messages
27
Likes
18
#2
Indium:

@senseless at AllMine experimented with Indium foil after we discussed and researched a few alternatives for TIMs. With the use of some Conformal Coating material (which is used partially as a TIM in its own right) to isolate the electrically conductive indium from exposed contacts, he was able to cover a majority of the unit in a combination of indium and other modest thermal pads (Pictured below).
Indum + Conformal Coating Thermal Interface Upgrade Rev1.jpg
(click to view full size)
Pushing 135W (at the wall) through the FPGA, on the stock air cooler but with the pictured upgrades, one unit achieves 56.6C and the other 62C. Prior to this upgrade, these units went to ~85C running the same bitstream on the same settings. This is a massive improvement, with two-fold benefits. Better controlled thermals means higher efficiency on power delivery, which means more easily controlled thermals on the power delivery system. If you can better manage the power delivery system performance, you can overclock the card higher, or run it harder in general.

Additionally, exposing the copper under the nickel- or zinc-plated stock VCU heatsink - by means of dremel or sanding - improves thermal transfer as well, given that you make make up for the loss of material by increasing your TIM thickness. This can drop your temps another 5-10C. You will also likely need to fill in the cross-hatch pattern with degassed thermal grease (or leave the existing grease there). In standard practice, this increases the effective surface area (by effective, I mean useful - the math works out to less paste between the sink and device overall, by sacrificing some surface area to purposely handle the worst of the problems in the paste) of the sink-to-device contact patch. My suspicion is that with something as high-performance as Indium (which has its own cross-hatching), the VCU's cross-hatched surface area (facilitated normally by only paste) performs worse than a perfectly flat contact patch that was 100% in contact with the indium. I will be able to confirm this in the coming months.
If you continue to use thermal grease or paste, then the cross-hatch pattern is necessary.


---
Update Sep. 21:
@senseless "Scavenged some larger tension springs and large m2 screws. I think the springs came from an old xeon v2 or v3 supermicro heatsink. FPGA temps dropped from 62-63c to 59-60c. LTC temp sensors 61c/68c. This is the board that was running a little hotter and pulling some extra amps."
Indum + Conformal Coating Thermal Interface Upgrade Rev1.1MountUpgrade.jpg
(click to view full size)
@senseless:
yep, 40Mh/s per card, but we have room to add additional cores and try to get better placement. I'd say we'd probably cap out at 80Mh/s per card with more work.
Later, senseless added a fairly large aluminum sink to the bank of the board around his new mounting solution.
Indum + Conformal Coating Thermal Interface Upgrade Rev1.2AluminumSink.jpg Indum + Conformal Coating Thermal Interface Upgrade Rev1.2AluminumSink2.jpg
(click to view full size)
 
Last edited:
Top Bottom