[DandyWorks] had an NVIDIA RTX 3070 Ti GPU, and found it was running incredibly hot, with the card’s memory hitting temperatures of 110 °C. He decided to try “copper modding” to solve the problem, and made some impressive improvements along the way.
Copper modding is where small copper shims are used to connect hot chips on the GPU to the heatsink more effectively than the standard thermal pads used by the manufacturer. Copper has much better thermal conductivity than thermal pads, and thus can help improve cooling of components when used in this fashion.
With the GPU carefully disassembled, [DandyWorks] notes the design uses a sub-heatsink specifically for the memory chips. He then sets about removing the thermal pads from the chips with isopropyl alcohol to help. They’re replaced with copper shims of a precise thickness, with a thin layer of thermal paste to ensure good heat flow. [DandyWorks] also shields all surrounding parts of the board with Kapton tape to avoid shorts if the copper shims happen to shift at any point.
Running the same hashing operation, the GPU now operates with its memory at a much cooler temperature of just 64 °C. [DandyWorks] ran the test for hours and temperatures didn’t climb beyond there. It’s evidence that the copper shims do a far better job of conducting the heat out of the memory chips versus the stock thermal pad setup.
We’ve seen some other interesting mods in this vein before, such as CPU die lapping for better thermal performance. Video after the break.
“with its memory at a much cooper temperature”
Copper on the mind? :P
Thanks, fixed.
Cooper Pairs occurs at low temperatures, so I for one would approve transforming this typo into a real idiom!
Why not recycle the heat, convert to usable Energy, and shunt it back to the motherboard?
Now that would be worthy of an HackaDay EE!
“Running the same hashing operation”
Oh, so it’s people like this guy who are why gamers, *and* game developers alike, are having trouble sourcing modern GPUs. Great of HaD to support these kinds of people.
HaD supports a widely understood Hacker community like the site name implies. Video mentions also other applications, like rendering, but the guy does indeed seems to be a crypto fan and uses hash computing as torture test – not sure how who he is relevant to this hack tho.
Criticizing someone for “misusing” consumer hardware on a hacking new site… lol
Time for some equal share of gpu resources! It can’t be that those who pay more for the product or secure it first, get it before common workin…gaming class, right commandre?
So why are gamers somehow “better” than miners? How does this ethical system work?
Because obviously wasting precious energy for only slight productive ends is more heinous than wasting precious energy for absolutely zero productive ends. ;-)
Didnt watch the video, but ‘hashing’ is a popular/default stress test for gpus (and cpus).
Edit: more like benchmark, not stress, but yeh, i dunno. (Disclaimer: i have zero blockchains)
Yeah, all those gamers with their Damn gamer gpu arrays!
I guess in your mind it’s more “productive” using 100x the expensive rare metals and fossil fuel energy for high-tech gambling than using one gpu to have an hour of fun after work.
is this the “oMG” cable cronie?
since its a separate heatsink, wouldnt it be possible to shorten the mounting studs, so it has “higher/actual” mounting pressure?
I look at the problem of extracting ~300 watts of heat and my mind jumps to increasing the mass of the heatsink (I suppose you could think of as increase the size of the thermal reservoir or making a larger thermal buffer between the heat source and the the heat destination)
300 watts is 300 joules of energy that needs to be dissipated every single second.
specific heat capacity of pure copper is 385 J/kg/°C ; density 8.96 kg/m³ ; thermal conductivity of 401 W/m/°C
specific heat capacity of Air (in a typical room) 1012 J/kg°C ; density 1.2 kg/m³ ; thermal conductivity of 0.026 W/m/°C
Looking at the above information the useful factors you have control over are the mass of the heatsink, surface area of the heatsink and the speed of the air being forced through the heatsink, and the volume of colder air (open doors vs closed doors if inside a building). There are fomulas but the simple rules would be to increase the mass and surface area until the desired operating temperature is reached. You can increase the fan speed but there is only so much air that you can force over the whole surface area before it becomes less efficient due to drag (friction).
Mass of the heatsink has no bearing at steady state. It is a side effect of larger chunk of metal that increase cross-section area – better spreading of hotspot, larger surface area for contact to air. Read up any heatsink design technical literature, you won’t find mass in the equations. (Been there)
It’ll help if/when the heating is low duty cycle AND the removal is below the peak. If your CPU/GPU load is near 100% and a few minutes, then it won’t do anything.
BTW The thermal equivalent circuit for thermal mass is a capacitor. As we all know that capacitor doesn’t help when you have a DC circuit after steady state. It’ll help to smooth out the peaks when you have AC (pulses of heat).
Tht’s definitely true, that said, current CPU can vary their power output so quickly, having a larger mass to dampen this can’t hurt…
I’ve done that sort of design when the heat is generated for a short period of time, such as supply discharge. It is also useful for averaging power, such as in switching power supplies. But for general design this only delays the length of time that the circuit will operate.
I hate being that guy but density 8.96 g/cm3 is closer to the mark.
Back in the good ol’ days, I had copper shims in my eeepc to bring down temps. It let me overclock that poor little atom chip way too high… good times.
5:43 Under first memory chip, is this resistor mounted diagonally or stray component from elsewhere?
I’m guessing it’s just mounted diagonally. There are other diagonally-mounted components visible.
One of the major uses of GPUs is to do parallel computing of scientific and engineering simulations for thing like… thermal transfer analysis. Kind of shocking that a company with the HPC chops of NVIDIA would specify rediculously thick 1.5mm thermal pads on a high end video card.
it’s not NVidia making the cards, just the chips. Cards are made by groups like EVGA who buy the chips from Nvidia and then build a card around it
Can i superglue the cooper shim directly to the heatsink? I mean heatsink, thermal paste and than superglue to the corners to keep the shims im place.
Only if you want more problems.
I have used old US (1984ish) copper pennies for shims after lapping them, attach with solder if possible. I’m too lazy to look up solder thermal properties, I just assume its better than pads or grease.
I’ve actually purchased “thermal glue” (by accident)- it’s the same as thermal paste but for permanent installations (came in 2 parts like epoxy).
The fact that to do all of this he also reseated the heatsink, cleaned the fans and reapplied thermal paste also helped a lot with temperature reducing.
“With the GPU carefully disassembled” I thought the GPU was the nVidia specific name for the accelerator IC placed on the graphics card?
Xerox a copy of your complaint while you cry in a kleenex and scotch tape it into an environmentally responsible used kraft envelope, mail it off, and I’m sure Hackaday will be on it like velcro.
hello, first of all thanks for this nice tip. can someone tell me how thick (what is the height-width-thickness in mm) of the copper plates? in short, all sizes have a weight.
What is the thickness of the plate to have on the VRM ?