← Back to Wiki
Hardware / GPU Repair

Replace the Thermal Pads on Any GPU

Memory that throttles, power draw that sags under load, a card that will not hold the hashrate it used to. Pads dry out and some ship bad from the factory. The card in this example went from throttling to holding 95 MH/s after a pad change.

Share on X

Find out what your card takes before you order

Search for your exact model and someone who has already been inside it.

What you want from them is the pad thickness, per area of the card.

Buy sheets rather than a kit. A 200 by 200 millimetre sheet is cheap.

Buy a range, from 0.5 millimetre up to 3 millimetre, if you work on more than one card. One card will often use three different thicknesses, sometimes five.

Get a pair of calipers.

Measure the old pads

Push the caliper down onto the old pad and read it.

It will not be exact, because a used pad has been squashed for two years.

It gets you close enough to pick the right sheet, which is all you need.

Match the thickness. Do not go thicker to be safe. A pad that is too thick holds the cooler up off everything near it. On a card covered in memory modules and capacitors, one over-thick pad means a row of components that are no longer touching anything. You will have made the cooling worse while believing you improved it.

Get the cooler off

Find the four screws with springs on them. Nearly every card has them.

Those four hold the heatsink down onto the GPU core, and they are the ones that matter on reassembly.

Take out the remaining screws holding the cooler to the card.

Note which screw came from where. They are not all the same length.

Lift the cooler slowly.

Watch for the fan headers. Bigger cards also have lighting or a display panel wired to the cooler. Unplug them rather than pulling until something gives.

Open the backplate too. This is the mistake worth avoiding. On many cards the backplate screws in from the component side, so you have to take the cooler off first and then remove the backplate separately. Skip it and you have done half the job. On the card in this example the backplate covered only three pads, and the memory modules it was supposed to be cooling were partly uncovered. That was the actual fault, and it took a second teardown to find.

Check what you took off

Look at the pads you removed.

Pads that have gone oily and are leaving a film behind have broken down. Replace all of them, not just the ones over the hot parts.

On the card here the 1 millimetre pads were fine and the 2 and 3 millimetre pads were oily. Thickness and batch both matter, so judge each one.

Put the pads back

Do the backplate side first, then fit the backplate.

Flip the card and do the core side.

Peel the film with a plastic tool from a repair kit rather than your fingers. Finger oil on a fresh pad defeats the point.

Keeping one colour per thickness makes it obvious later which pad went where.

Paste and reassemble

Use a stable paste rather than the best performing one.

MX-4 holds up for years. The high end pastes score better on day one and dry out, which is the wrong trade for a card that runs 24 hours a day.

Spread a thin even film with a spreader instead of placing a dot. Mounting a heavy cooler blind is fiddly enough without hoping a blob spread correctly.

Reconnect every fan header and anything else you unplugged.

Start all the screws before you tighten any of them, so the cooler sits square.

Tighten the four spring screws in a crisscross pattern, a little at a time, corner to opposite corner.

Stop when the screw stops. Do not lean on it.

Then do the rest of the screws.

Test it under real load

Put the card back in and check temperatures at idle first.

Then run it hard for 10 to 20 minutes and watch the power draw at the wall.

Power sagging under sustained load is the signature of the card pulling itself back, and it is what a pad job fixes.

Once it holds, step the memory clock up gradually and find where it stops being stable.

The example card held 95 MH/s afterward, against about 70 while it was throttling and cooking its own memory.

Weigh this against a warranty claim. Opening the card is your call to make, and on a current card under warranty the claim is usually the better move. The reason to repair anyway is supply. A vendor with nothing to send you may simply refund you, and then you have money and no GPU. Decide which of those you would rather be holding before you pick up a screwdriver.