Nvidia was recently engulfed in the melting 16-pin connector fiasco, but now things are heating up for AMD. Reports and user feedback have shown that the Radeon RX 7900 XTX can get sizzling hot, with hotspots hitting 110 degrees Celsius. As a result, PowerColor, one of AMD’s notable partners, has issued a call for user reports on Reddit (opens in new tab) so it can send the details to AMD for further investigation.
Thus far, the 110C hotspot phenomenon appears only to affect MBA (Made By AMD) or reference Radeon RX 7900-series graphics cards, whether you purchased them directly from AMD or one of the chipmaker’s partners. There haven’t been any reports of custom designs afflicted with the issue, suggesting a potential design flaw with the reference cooler. However, let’s not get ahead of ourselves: The million-dollar question is whether the 110 degrees Celsius junction temperature is within specification for Radeon RX 7900-series graphics cards. We’ve reached out to AMD for confirmation; however, the chipmaker hasn’t gotten back to us yet, probably due to the holidays.
There’s a precedent of Radeon graphics cards reaching 110C, which aligns with AMD’s specifications. If you don’t remember, the normal junction temperature for the Radeon RX 5700-series graphics cards, such as the Radeon RX 5700 XT, was 110C.
“Instead of setting a conservative, ‘worst case’ throttling temperature for the entire die, the Radeon RX 5700 series GPUs will continue opportunistically and aggressively ramp clocks until any of the many available sensors hits the ‘hotspot’ or ‘Junction’ temperature of 110 degrees Celsius. Operating at up to 110C Junction Temperature during typical gaming usage is expected and within spec,” stated AMD in a blog post (opens in new tab) three years ago.
Besides the junction temperature, it’s essential to look at the edge temperature to determine whether there is an underlying problem. The edge temperature corresponds to the temperature measured at the edge of the silicon. So it’s an excellent metric to reveal the condition of the mounting pressure. A low value means there’s good mounting pressure. “110 hotspot is within spec if your card is getting up to 90° GPU edge temp... If 110 is hit at 70 edge, yeah that is not ideal,” commented an AMD representative on Reddit (opens in new tab).
AMD has already said that its GPU team is investigating the problem. It’ll be essential to evaluate if the high temperatures are the culprit for the thermal throttling, forcing the graphics card to fall below the rated clock speeds. Otherwise, it’s normal behavior. Remember that AMD only guarantees the base clock speeds for its graphics cards.
Meanwhile, owners are reportedly RMAing or returning their Radeon RX 7900 XTX graphics cards. The feedback has been a mixed bag, though. In a Reddit thread (opens in new tab), some said AMD had accepted their RMA, while others claimed the chipmaker rejected the RMA and refunds.
Some user reports say repasting solves the problem on the Radeon RX 7900 XTX, but it’s not a 100% fix. There is a theory (opens in new tab) that the cooler’s cold plate is uneven. That would explain why the thermal interface material (TIM) looks thicker in the middle than on the edge of the die. One Redditor (opens in new tab) suggested that a graphite thermal pad could solve the problem. That’s interesting because AMD opted to use thermal pads in previous generations of Radeon graphics cards, so it’s unknown why the chipmaker went for TIM on RDNA 3. Either way, we don’t recommend you tear off your cooler since it would likely void your warranty. The only option right now is to wait for AMD’s guidance or just attempt to RMA the graphics card, though the latter would be a fruitless endeavor if the card proves to be operating as expected.
At least thats what Der8auer find out after putting some "new", new thermal paste on his 7900XTX.
A real pitty if this really issue, and a real turn down for all the people who got this card and may not want to mess up with the heatsink cause it probably voids the warranty.
That is incorrect. You only lose the warranty if you break or damage the card by doing so. I would agree that end users shouldn't need to in a friggen $1K device, unless they want to, but saying you'd lose the warranty is misleading and a tad fearmonging.
Still, AMD better get some answers soon. If it is a problem with the cooler, they better start making good on users that went and supported them with the reference cards on day 1. As a side note: this is why I wait for AIB models (specially Sapphire) on the AMD camp: gives people more time to "field test them" and AIB models usually are a tad more refined, cooling-wise.
Or is it by feel or machine?
Saldy I have no idea who make this batch of cards for AMD. Im guessing they were in a rush to get the cards out and into the world in time for day 1.
Im guessing if all this issues and news cover are because of a screw torque was not enough, then AMD wont be very happy with them lol
In the past, Sapphire has been the OEM AIB for reference models. I wonder if they switched suppliers because Sapphire is known for cooling and robust VRMs, something this reference design is lacking.
The screw assembly drivers are pre-torqued. Most are done by hand in the finishing stages of board assembly. Things like cables are hand inserted. As these are attached to the HS, they are usually done at the same workstation. (Based on the videos I have watched) But I have personally seen factories where this process is automated. Cooler gets hand loaded into a jig and the holes aligned by machine. Then the screw setters automatically screw them in till torque is reached.
There also seems to be an issue with vapor chamber designs running hot. (Same as 5700XT, but not 6000 series) If I were to guess, I would say they need a good lapping to even them out...but money. Over tightening the screw may apply enough pressure to the malleable heat plate to cause it to flatten out. But it's risky. Stripping threads, or cracking a chiplette, board is definitely a risk.
Radeon 5970 Overclocking: The VRM Temperature Bottleneck (anandtech.com)