4 ms·

Heat pipes - the large copper tubes between hot chips and heatsinks - are actually hollow and usually filled with some liquid/gas mixture. That's often butane

Heat pipes - the large copper tubes between hot chips and heatsinks - are actually hollow and usually filled with some liquid/gas mixture. That's often butane or something with a lowish boiling temperature at lowish pressures. They work by having the liquid boil at the hot end, the gas flow along the pipe, condense back to a liquid at the cold end, and then the liquid flow back. That last step, the liquid flowing back, is the critical one. With gravity, the liquid can gush back really fast, and loads of heat can be transferred. But if the card is the wrong way up, then it relies on the liquid wicking back along some clothlike material coating the inside of the pipe. That's a far slower process. Therefore the heat that can be dissipated this way is much lower. That's why heat pipe performance is dramatically affected by orientation. The designers of this card will have been aware of that, but likely decided a small performance hit for sustained workloads was probably worth it.

onli 4y ago
It's not a small performance hit. The card is throttling a lot and ramping up the fans up to 1000RPM more than usual because cooling fails completely. There were report about cards failing while testing this (e.g. in the der8auer video), but ofc it's not possible to be sure that the card broke because of this. It's possible though. This is not an expected failure mode the engineers would have been okay with even if it did not kill cards. No one would be okay with the heatpipe drying up like this.
- formerly_proven 4y ago
  The card der8auer had which failed was very clearly a VRM failure.
  - onli 4y ago
    That's not very re-assuring when the cooler is not working as expected. VRM gets killed by heat (I don't actually think it's likely that this is what happens here, but I do think AMD is making a mistake by letting customers run into this issue and by giving me reason to speculate like this).
    - tgtweak 4y ago
      I mean no VRM or vmem is meant to run at 110C sustained. The 3090 design had an issue with vram cooking leading to 105+ temps and the gpu would throttle when it got there. Many people that didn't mitigate that problem (with extra cooking) experienced memory failure. The main problem was vram temps aren't recorded so it looked like the GPU temps were fine. Source: everyone who mined Ethereum on a 3090.
      - runnerup 4y ago
        > everyone who mined Ethereum on a 3090. Most 3090's used for ethereum mining used reduced power load and temperatures. https://www.reddit.com/r/EtherMining/comments/q2v061/comment/hfnslen/?utm_source=reddit&utm_medium=web2x&context=3 https://www.reddit.com/r/EtherMining/comments/q2v061/comment... e.g. this guy: > That’s pretty terrible efficiency. I can get 120 with only 290W. For me the extra 4MH/s does not pay for 60W. Some individual enthusiasts were trying to get every last "megahash" for benchmarking bragging rights, but anyone who had a bunch of 3090's would minimize power consumption because the marginal profit on the last 50+ watts was negative, which had the nice side effect of lowering temperatures.
        code_biologist 4y ago
        +1. I mined on a few 3060s (and heated my house) last winter. All with cores underclocked, memory overclocked, all at ~60C temperature wise, each consuming about 120W.
        tgtweak 4y ago
        3090 on ethereum (even after undervolting/underclocking gpu and bumping vram frequency) was hitting 100+C on VRAM. You went from 95MH/s to 120MH/s by doing this. Throttling happened at 110C so you wouldn't notice it on hashrate unless you had really poor cooling. It was causing premature failure on vram even in the first months of use at these temperatures. If you put some extra heat syncs on the card and got vmem down to ~90C you had no issues. Almost all miners were doing this. If you didn't cut back the gpu voltage/clock it would increase vram temps to 100+ even at stock vmem frequency. GPU core temps dropped by 25-30C when underclocking/undervolting and power would also go down by 60-80W.
wmf 4y ago
The thing is that the vapor chamber is working perfectly fine upside down in most of these cards. It sounds like a manufacturing problem not a design flaw.
Dylan16807 4y ago
Well, the card actually performs better in the orientation where half the vapor is going down and wicking back up, and worse in the orientation where it's flat. But the bigger issue here is not the way orientation affects performance, but that defective cards seem to be hitting dryout. At that point a combination of not-enough-liquid and hot wick means that the liquid isn't making it back to the heat source before it boils again, and performance plummets.
spuz 4y ago
AMD have stated that the problem is a lack of water in the vapour chamber of a small batch of "made by AMD" 7900 XTXs. https://youtu.be/X87OzJ3bU7o?t=209 https://youtu.be/X87OzJ3bU7o?t=209
smolder 4y ago
Anecdotally, my friend got one he had to return because it maxed out the fans, throttled, and crashed, regardless of orientation. These are defective cards, and there are a lot of them. Don't explain away broken cards by calling it some design compromise. A high percentage of the AMD coolers are completely unusable due to defect, and should be replaced. The ones that work in one orientation but not the other are also defective. No one should have to put up with any dry-out whatsoever, in either orientation.

onli 4y ago

It's not a small performance hit. The card is throttling a lot and ramping up the fans up to 1000RPM more than usual because cooling fails completely. There were report about cards failing while testing this (e.g. in the der8auer video), but ofc it's not possible to be sure that the card broke because of this. It's possible though. This is not an expected failure mode the engineers would have been okay with even if it did not kill cards. No one would be okay with the heatpipe drying up like this.

formerly_proven 4y ago
The card der8auer had which failed was very clearly a VRM failure.
- onli 4y ago
  That's not very re-assuring when the cooler is not working as expected. VRM gets killed by heat (I don't actually think it's likely that this is what happens here, but I do think AMD is making a mistake by letting customers run into this issue and by giving me reason to speculate like this).
  - tgtweak 4y ago
    I mean no VRM or vmem is meant to run at 110C sustained. The 3090 design had an issue with vram cooking leading to 105+ temps and the gpu would throttle when it got there. Many people that didn't mitigate that problem (with extra cooking) experienced memory failure. The main problem was vram temps aren't recorded so it looked like the GPU temps were fine. Source: everyone who mined Ethereum on a 3090.
    - runnerup 4y ago
      > everyone who mined Ethereum on a 3090. Most 3090's used for ethereum mining used reduced power load and temperatures. https://www.reddit.com/r/EtherMining/comments/q2v061/comment/hfnslen/?utm_source=reddit&utm_medium=web2x&context=3 https://www.reddit.com/r/EtherMining/comments/q2v061/comment... e.g. this guy: > That’s pretty terrible efficiency. I can get 120 with only 290W. For me the extra 4MH/s does not pay for 60W. Some individual enthusiasts were trying to get every last "megahash" for benchmarking bragging rights, but anyone who had a bunch of 3090's would minimize power consumption because the marginal profit on the last 50+ watts was negative, which had the nice side effect of lowering temperatures.
      - code_biologist 4y ago
        +1. I mined on a few 3060s (and heated my house) last winter. All with cores underclocked, memory overclocked, all at ~60C temperature wise, each consuming about 120W.
      - tgtweak 4y ago
        3090 on ethereum (even after undervolting/underclocking gpu and bumping vram frequency) was hitting 100+C on VRAM. You went from 95MH/s to 120MH/s by doing this. Throttling happened at 110C so you wouldn't notice it on hashrate unless you had really poor cooling. It was causing premature failure on vram even in the first months of use at these temperatures. If you put some extra heat syncs on the card and got vmem down to ~90C you had no issues. Almost all miners were doing this. If you didn't cut back the gpu voltage/clock it would increase vram temps to 100+ even at stock vmem frequency. GPU core temps dropped by 25-30C when underclocking/undervolting and power would also go down by 60-80W.

wmf 4y ago

The thing is that the vapor chamber is working perfectly fine upside down in most of these cards. It sounds like a manufacturing problem not a design flaw.

Dylan16807 4y ago

Well, the card actually performs better in the orientation where half the vapor is going down and wicking back up, and worse in the orientation where it's flat. But the bigger issue here is not the way orientation affects performance, but that defective cards seem to be hitting dryout. At that point a combination of not-enough-liquid and hot wick means that the liquid isn't making it back to the heat source before it boils again, and performance plummets.

spuz 4y ago

AMD have stated that the problem is a lack of water in the vapour chamber of a small batch of "made by AMD" 7900 XTXs. https://youtu.be/X87OzJ3bU7o?t=209 https://youtu.be/X87OzJ3bU7o?t=209

smolder 4y ago

Anecdotally, my friend got one he had to return because it maxed out the fans, throttled, and crashed, regardless of orientation. These are defective cards, and there are a lot of them. Don't explain away broken cards by calling it some design compromise. A high percentage of the AMD coolers are completely unusable due to defect, and should be replaced. The ones that work in one orientation but not the other are also defective. No one should have to put up with any dry-out whatsoever, in either orientation.