The attachment of the heatsink is quite questionable
Since AMD works with four threaded sleeves and a clamping cross, you can tighten the screws as much as you want – the torque will not change after a certain moment. However, since all sleeves are the same length (and thus the torque on all screws is always identical), something like this only works if the heatsink underneath is really absolutely flat. If not, the contact pressures differ extremely, even with identically tightened screws! Exactly that with the straight support, however, was not the case with my copy. This could then also explain why in some cases replacing the thermal paste does virtually nothing and the temperature gradient between GPU temperature and hotspot remains identical.
Compared to NVIDIA, where the cooler block including vapor chamber and its heatsink remains independent and flexible from the rest of the board mounting, AMD unfortunately relies on an almost monolithic construct instead of a loose insert. This means that the heatsink cannot adjust to any gaps or inaccuracies independently of the rest of the structure, but is firmly fixed. As long as the horizontal alignment is parallel and flat to the GPU in all planes, there is nothing wrong with it. When…
If you now put the board on it, you can see that the sleeve even still clearly protrudes! Together with the limitations of the spring cross, each of the four screws can therefore only be screwed on tightly up to an identical and limited value. If the heatsink or the GPU (or even both) are not really flat and also have height differences at all possible points, then the resulting contact pressure is naturally very different at all points. We’ll keep that in mind for once, because now we’re measuring!
GPU vs. heatsink – nothing fits together there
The GPU is slightly convex, which is not a big deal. You will never get it completely straight anyway, and those who remember the “humpback whales” in NVIDIA’s first Samsung GPUs and my analyses will of course know what I mean by that. At that time, it was probably the inappropriately executed lamination of the GPU board (shortcoming of the original foil), which shrank slightly and even warped the die along with it. In contrast, the values measured here are still really tolerable. The measured 0.07 mm difference between edge (chiplets) and center (GFX) is nothing that couldn’t be easily compensated with a good gap filler. Especially since these differences turn out to be almost identical at all corners. Thus, the GPU is ruled out as an uneven surface destroyer for now.
Which brings us back to the heatsink. This time I used the maximum resolution of the scanner, which unfortunately also took two days for this small crop, but it was really worth it. Let’s take a look at the height differences here as well. From left to right, i.e. lengthwise to the GPU, we see a slight increase of 0.04 mm on the outside and in the center. Thus, the heatsink is not quite at the same angle to the GPU, but defines a slightly inclined plane. In addition, it is also curved differently between the narrow sides.
Considered individually, these are also normal and quite acceptable tolerances. However, if you put the “inclined” and once again slightly curved plane on the already convex surface of the GPU, you end up with a difference of up to 0.11 mm. that is rather ugly but also still nothing that you could not somehow compensate with a good paste. But if you look at the mounting, then the best paste is actually useless if the contact pressure is partially insufficient because the screw connection does not provide more.
Summary and conclusion
Within such tolerance ranges, it is certainly possible to bypass the stoppers, use suitable screws with larger heads and neutralize the protruding sleeves by means of higher washers, so that when screwing with new, good paste, the contact pressure can be better dosed. Since I’m not AMD and also don’t know the maximum permissible values, I’ll leave a question mark here for safety reasons. Bending hollow chip surfaces with force just to make them fit the crooked and curved heatsink is certainly associated with a certain risk. This, however, is not something I want to impose irresponsibly on any of my readers. Everyone must then decide for themselves.
However, with a suitable thermal paste (Alphacool Apex B-Stock, so the firmer), two washers per hole and suitable screws instead of the clamping cross, I could also lower the hotspot that occurred with me even compared to the “normal” card to a slightly lower delta of 15 to 17 Kelvin, which has not worked by simply changing the paste. I take this for me personally as a conclusion and still explicitly refer to the preface on page one. It was the comprehensible and also measurable reason for my card, but it does not necessarily have to be for other cards. It can also be directly due to the chamber and possible production or construction errors, the last word is certainly not yet spoken. But you need much more time and samples for that.
It is certainly also conceivable that individual sensors in the GPU measure mischief and the telemetry does not catch such outliers (plausibility test). This could also be solved by AMD via a firmware update. However, today’s test speaks against this (at least for me), where the temperatures of the six chiplets suddenly turned out to be almost the same again after the reassembly without the clamping cross, where there was still up to 8 Kelvin difference before and my card even performed better than the original without the hotspot problem. In any case, it is advisable to wait for a statement from AMD. I for one have at least been able to solve my problem.
By the way, I also had the slight difference between vertical and horizontal mounting with the “repaired” card and two other, rather unsuspicious models, which could clearly be blamed on the vapor chamber and cooler design as such. But I don’t think that had anything directly to do with the hotspot, which is why it’s actually a different topic. If anyone is experimenting around with their RX 7900 XT(X) or still finding problems, I would be very happy to hear from you. In the meantime, I’ll wait and see what the swarm knowledge and AMD bring to light (or not).
371 Antworten
Kommentar
Lade neue Kommentare
Urgestein
Mitglied
1
Veteran
1
Veteran
Mitglied
1
Urgestein
Mitglied
Urgestein
Veteran
Urgestein
Veteran
Urgestein
Urgestein
Urgestein
Alle Kommentare lesen unter igor´sLAB Community →