Thermal interface: paste, pads and phase-change

The GA100 die and its HBM2e stacks on the CMP 170HX sit under an integrated heat spreader (IHS), so the card has two thermal interfaces in series: a factory one between the die and the IHS, held under a glued lid, and the one between the IHS and the cooler, which is what you replace when you repaste. Around the package, pads (or putty) carry heat from the power components and the small parts near the core into the heatsink. On used cards the paste on top of the IHS is often dry and pads are missing, but a fresh interface on top gains anywhere from nothing to a large drop, and many owners blame the old paste under the lid. Opening and cleaning the card are on Disassembly and cleaning, temperature limits on Thermal limits and water blocks on Water cooling.

Repasting on top of the IHS

Repaste results range from no change to 85 to 58 C

Contested 8 GB and 10 GB Results of a fresh interface between the IHS and the stock cooler, by side:

Gains: @ju5t5omedude (2026-07-31) relayed another user's before/after chart for a card cleaned and fitted with new paste and pads, without delidding, on a card serviced for resale: 85 to 58 (C presumably). On 2026-08-08 he pointed to before/after charts by Alex1997 in another server's cooling section showing a significant improvement. @roger_dammit found that sealing the open areas of his case also helped a lot. @ppc64_ (2026-08-11) found the stock paste quite dried out; after a repaste with 2 mm pads, adding pads to heatsink spots that had none, the HBM ran well under 10 C above the core. @ghostinthedatacenter_87281's repaste and repad brought memory 13 C lower on a 10 GB card. Source: @ju5t5omedude, 2026-07-31
Source: @roger_dammit, 2026-07-31
Source: @ju5t5omedude, 2026-08-08
Source: @ppc64_, 2026-08-11
Source: @ghostinthedatacenter_87281, 2026-10-20
Little or no gain: @danske1 asked @meromgreen whether his repad made a difference; @meromgreen (2026-08-08) saw no noticeable change, maybe 1 C lower maximum, and thinks his Arctic P12 Pro PST fans are the limit. @rogersmith2026's core temperature dropped less than he hoped after a 2 mm repad. @catchyuser still saw 85 to 87 C at 100% GPU utilization after an Arctic MX-7 repaste (2026-08-22). @finishtilt14 saw little difference from a repaste and repad years ago. @alanoo repasted one card with MX-7 that week and gained absolutely zero (2026-10-16); @a1rm4x took part in that thread. Source: @danske1, 2026-07-31
Source: @danske1, 2026-08-08
Source: @meromgreen, 2026-08-08
Source: @rogersmith2026, 2026-08-11
Source: @catchyuser, 2026-08-22
Source: @finishtilt14, 2026-09-16
Source: @alanoo, 2026-10-16
Source: @a1rm4x, 2026-10-16
No change from any paste: @skirtchaser tried every paste he had on top of the IHS, from old Arctic Silver 5 to Thermal Grizzly Kryonaut, with no temperature change (2026-09-07; see #Paste or solder under the IHS). Source: @skirtchaser, 2026-09-07

The largest measured gain in the history came from delidding, not from repasting: see #Does delidding help?.

When a repaste is worth it

Contested 8 GB and 10 GB Whether to open a card that runs acceptably, in the order posted:

@rubbra_f and @sealclubber (2026-07-29): clean and repaste with a generous amount of non-conductive paste, because the dies are large. In the same thread @danske1 bolted 140 mm Noctua fans on as a stopgap, @deejay9983 used printed shrouds with 120 mm fans and @jimtjames waited for water-cooling parts because printed shrouds did not fit several cards in his server. Source: @danske1, 2026-07-29
Source: @rubbra_f, 2026-07-29
Source: @sealclubber, 2026-07-29
Source: @deejay9983, 2026-07-29
Source: @jimtjames, 2026-07-29
@_arcanum_ (2026-08-08) had not repasted his own cards; one buyer had asked him for a repaste. Source: @_arcanum_, 2026-08-08
@irythros (2026-08-22, answering @catchyuser's hot MX-7 card): good contact on both the die and the memory to the cooler is required, and the TIM under the lid may have dried out; one photo in the server showed a cooler ripping the lid off with cooked TIM. He recommends PTM7950 (see #PTM7950: price, counterfeits and handling). Source: @catchyuser, 2026-08-22
Source: @irythros, 2026-08-22
@schakaqt (2026-09-10) did not repaste because temperatures were already fine and the cards have an IHS over the die; @cibernox, who had asked whether cards a seller had already cap-modded need opening, decided to stress test first. Source: @cibernox, 2026-09-10
Source: @schakaqt, 2026-09-10
@_orrorin (2026-10-12): these are mining cards and the paste may be bone dry; check the card's hours used and repaste if they run into the thousands (his nearly new 10 GB cards were not repasted). @vcore2191: if temperatures are fine under long load at the fan speed you will run, a repaste typically gains about 2 C unless the old job was bad. @ob88594: generally unnecessary, disassembly is very prone to damage, and for a real drop go liquid, about $200 more. @alanoo: cards bought already cap-modded were probably repasted by the modder. @supermanisdum took part in the thread. Source: @_orrorin, 2026-10-12
Source: @alanoo, 2026-10-12
Source: @vcore2191, 2026-10-12
Source: @ob88594, 2026-10-12
Source: @supermanisdum, 2026-10-12
@pekoraali (2026-10-19) found dried paste on two of three cards but avoids repasting after knocking an SMD capacitor off a 5700 XT. Source: @pekoraali, 2026-10-19
@civiloid (2026-10-20): do not overtighten the stock heatsink. Source: @civiloid, 2026-10-20
@segmond (2026-10-22) asked whether a hot-running card is safe to repaste yourself; the answer was yes if you can follow instructions, and the supplies are cheap. Source: @segmond, 2026-10-22

Dried paste or starved airflow: reading core against HBM

On the lab's eight-card server, the hottest cards ran core and HBM within 3 C of each other, which points to starved airflow rather than dried paste. Dried paste shows as a core far above its own HBM. Check this before you open a card that runs hot in a crowded chassis; the airflow side is on Fans and ducts.

Reported 8 GB
Source: @darkmaniac7, 2026-10-28 (lab notes)

Paste choice for the large IHS

Contested 8 GB and 10 GB The IHS is large, so it takes a lot of paste; which paste suits it, in the order posted:

@quark2world (2026-08-12): apply paste in an X pattern so the pressure spreads it, and buy a large tube because the IHS is huge. @skirtchaser recommends GD900 paste, and @_studebaker_ uses Maxtor MTP-8301C (150 g, 12 W/mK). Source: @quark2world, 2026-08-12
Source: @skirtchaser, 2026-08-12
Source: @_studebaker_, 2026-08-12
@cute_sc2 (2026-09-01) used Thermal Grizzly pads and paste for a full repaste of every card. Source: @cute_sc2, 2026-09-01
For the top layer, @_arcanum_ uses Shin-Etsu 7921 and calls phase-change material the lowest-maintenance but pricey option, while @skirtchaser picks Thermal Grizzly Carbonaut as the most set-and-forget (2026-09-07). @rrickk420 and @chonsworth took part in the thread. Source: @_arcanum_, 2026-09-07
Source: @skirtchaser, 2026-09-07
Source: @rrickk420, 2026-09-07
Source: @chonsworth, 2026-09-07
@_orrorin (2026-10-12) recommends Arctic MX-7 (thick, long-lasting, high conductance) and a lot of it (he exaggerated "8 grams"); @alanoo calls MX-7 the worst choice for a big surface because it is hard to spread, and notes it looks dry even when new. Source: @_orrorin, 2026-10-12
Source: @alanoo, 2026-10-12

Phase-change material on the IHS has its own section below.

Video guide: "Nvidia CMP 170HX: The Repastening"

Asked for a write-up on replacing the thermal pads, @globalapathy pointed to the YouTube video "Nvidia CMP 170HX: The Repastening", a couple of years old, from a mining channel. A comment under it gives the pads as 1.5 mm; the pad readings are in #Thermal pads.

Reported 8 GB and 10 GB
Source: @fiiireface, 2026-10-02
Source: @globalapathy, 2026-10-02

Under the IHS

Paste or solder under the IHS

Contested 8 GB and 10 GB What sits between the die and the IHS, in the order posted:

@farawaythrower asked on 2026-07-27 what the interface is and whether delidding to replace it is worthwhile; no answer came then. Source: @farawaythrower, 2026-07-27
@meatsus, after a failed delid attempt (2026-08-01): if it were tin/indium solder there would be no reason to delid. On 2026-08-03 he cited Google Gemini saying the A100 40GB uses indium solder, an AI answer rather than a teardown. Source: @meatsus, 2026-08-01
Source: @meatsus, 2026-08-03
@meatsus (2026-08-10): the IHS is soldered with tin-indium, not pasted, and heating the solder to its melting point on a package this large can destroy the chip, so do not try to remove it. He said people who delidded 40 GB A100s confirmed a soldered IHS and that the CMP must be the same, said this settled an earlier claim, repeated by others including @scallion_9883, that delidding helps because of dried paste under the lid, a claim he said he had started himself, and said repasting only applies between the IHS and the heatsink. @quark2world, @ju5t5omedude and @skirtchaser took part in the thread. Source: @meatsus, 2026-08-10
Source: @quark2world, 2026-08-10
Source: @ju5t5omedude, 2026-08-10
Source: @skirtchaser, 2026-08-10
Source: @scallion_9883, 2026-08-10
@meromgreen (2026-08-21) linked an arca.live post where a CMP's IHS appeared to come off with the heatsink. @_arcanum_: you cannot solder onto HBM because its degradation temperature is too low. @danske1 found the photo odd and wondered whether it was AI-generated; @meromgreen reopened his own card and showed the IHS and the heatsink are separate pieces of metal, the oily look being old pad grease. @meatsus warned that working around the package risks knocking off SMDs. @meme_meme took part in the thread. Source: @meromgreen, 2026-08-21
Source: @_arcanum_, 2026-08-21
Source: @danske1, 2026-08-21
Source: @meme_meme, 2026-08-21
Source: @meatsus, 2026-08-21
Source: @skirtchaser, 2026-08-21
The same day @quark2world said the IHS is held by the perimeter glue, not by the TIM, and the TIM under it is paste; @meatsus said LLMs had wrongly told him it was tin-indium solder and that the paste is about 6 years old; @dalzyyy found the glue rock hard with a razor; @meromgreen linked an arca.live post by another user who popped the IHS. Source: @meromgreen, 2026-08-21
Source: @quark2world, 2026-08-21
Source: @dalzyyy, 2026-08-21
Source: @meatsus, 2026-08-21
@skirtchaser (2026-08-27) posted a photo suggesting paste under the IHS; @alanoo: that would explain why repasting the cooler alone does not help much. @_arcanum_, @civiloid, @dalzyyy and @danske1 took part in the thread. Source: @_arcanum_, 2026-08-27
Source: @skirtchaser, 2026-08-27
Source: @civiloid, 2026-08-27
Source: @alanoo, 2026-08-27
Source: @dalzyyy, 2026-08-27
Source: @danske1, 2026-08-27
@skirtchaser (2026-09-07), after his top-paste test showed no change: the bottleneck is paste under the IHS rather than solder, since with solder the top paste would matter more. @ju5t5omedude: several people have delidded cards by accident while removing the heatsink, which also points to no solder. Source: @skirtchaser, 2026-09-07
Source: @ju5t5omedude, 2026-09-07
@meatsus's deliberate delid of 2026-09-15 found the material under the lid hard and rubber-like; his manual of 2026-09-19 speaks of 7-year-old paste underneath. Source: @meatsus, 2026-09-15
Source: @meatsus, 2026-09-19
@spectakkular (2026-10-12): there is paste under the IHS too, which limits the ceiling. @wakeupmbrc1: someone already delidded and found paste, not solder. @globalapathy: normally the IHS-to-die interface is indium solder. Source: @spectakkular, 2026-10-12
Source: @wakeupmbrc1, 2026-10-12
Source: @globalapathy, 2026-10-12
@gnasherrr (2026-10-18): the interface is paste, apparently a phase-change material, not indium solder. @skirtchaser posted a photo of an accidental delid. @globalapathy believes the lid is held by silicone glue plus phase-change material and could be reattached. Source: @gnasherrr, 2026-10-18
Source: @skirtchaser, 2026-10-18
Source: @globalapathy, 2026-10-18
Source: @gnasherrr, 2026-10-20
How fast the paste under the IHS ages: @skirtchaser says it can degrade within months, citing the GTX 480; @endsjeremy doubts it. Source: @skirtchaser, 2026-10-18
Source: @endsjeremy, 2026-10-18

Does delidding help?

Contested 8 GB and 10 GB Whether removing the IHS and its old paste is worth the risk, in the order posted:

@cute_sc2 (2026-09-13) keeps stock memory speed because the paste between the die and the heat spreader on these cards is dry. Source: @cute_sc2, 2026-09-13
@meatsus (2026-09-15) compared setups on the stock cooler at a 150 W cap with one Arctic S8038 fan at 80%: original paste about 65 C core and 75 C memory; repasted on top of the IHS 60 to 65 C and 70 to 75 C; IHS removed, cooler on the bare die, 55 to 60 C and 60 to 65 C. Under the Vulkan memtest the memory fell from 72 to 74 C to 62 to 64 C. On 2026-09-21 he repeated about 10 C against old paste plus IHS: delidded, 150 W, one Arctic S12038-8K at 50%, HBM about 70 C and core about 60 C; for comparison @globalapathy (stock paste, a 140 mm fan at 100%) saw about 67 C core at 150 W in a memtest, and @okrasit. saw a spike at the start of load and planned to repaste. Source: @meatsus, 2026-09-15
Source: @meatsus, 2026-09-21
Source: @globalapathy, 2026-09-21
Source: @okrasit., 2026-09-21
@finishtilt14 (2026-09-16): a delid mattered more than his own repaste and repad, about 10 C. Source: @finishtilt14, 2026-09-16
@wakeupmbrc1 (2026-09-19): a delidded card can use thinner pads on the other components with the same cooler. On 2026-10-12 he said delidding will not help, because the die is huge and thermal resistance matters less. Source: @wakeupmbrc1, 2026-09-19
Source: @wakeupmbrc1, 2026-10-12
@xanthefin (2026-10-18): most of the thermal issue is that the paste under the IHS is old. Source: @xanthefin, 2026-10-18

Delidding: accidental delids, a cut board and the bare-die trap

Warning Removing the IHS can damage the card, and it also happens by accident while the cooler comes off. When @meatsus slipped a craft knife between the IHS and the die it felt like soldered Opterons; he stopped in time and the GPU still works, but he cut off a PCB corner (2026-08-10). @skirtchaser (2026-09-04): delidding seems possible but is an extreme risk. Several cards have been delidded by accident while the heatsink was removed (@ju5t5omedude), and @skirtchaser posted a photo of one; on 2026-09-09 @cute_sc2 and @scallion_9883 told @wakeupmbrc1 that a covered GA100 had been delidded before, somewhere in the Discord history. @civiloid: the stock cooler cannot go on the bare die without making up for the lid thickness ("you would cook the die"), a thick pad in its place is worse than reinstalling the lid, and the stock heatsink cannot be tightened further without damaging the VRM; bare-die blocks for the A100 40GB and 80GB PCIe are different parts. @xeax: delid plus a bare-die AIO is the only sensible path. @meatsus's way of making up the 0.5 mm IHS after a delid: 1.5 mm pads on the other components instead of the stock 2 mm, because thicker pads would bend the board, and 0.8 mm washers for the larger bracket gap; the core itself takes paste, not a pad (@c133 suggested Thermal Grizzly KryoSheet). Delidding is also what lets A100 80 GB water blocks fit; see Water cooling.

8 GB and 10 GB
Source: @meatsus, 2026-08-10
Source: @skirtchaser, 2026-09-04
Source: @ju5t5omedude, 2026-09-07
Source: @wakeupmbrc1, 2026-09-09
Source: @cute_sc2, 2026-09-09
Source: @scallion_9883, 2026-09-09
Source: @meatsus, 2026-09-15
Source: @c133, 2026-09-15
Source: @skirtchaser, 2026-10-18
Source: @civiloid, 2026-10-18
Source: @xeax, 2026-10-18

A longer write-up by @meatsus exists (170hx_delid_manual.pdf, 2026-09-19); see Discord. It waits for the author's yes before the wiki reproduces it.

Idea: delid, solder a copper shim to the stock cooler, phase-change material on the bare die

@markshowa proposes delidding the card, soft-soldering a copper shim onto the original copper cooler to remove one bad thermal interface, and putting a good phase-change material between the die and HBM and the copper. The goal is lower HBM temperatures, to clock the memory higher; the concern is the die pressure the IHS takes today. The idea needed the IHS thickness, an open question when he raised it, and @meatsus has since measured 0.5 mm. The soldered shim itself is untested.

Theory 8 GB and 10 GB
Source: @meatsus, 2026-09-15
Source: @markshowa, 2026-10-18
Source: @markshowa, 2026-10-19

Thermal pads

Stock pad thickness: caliper readings disagree, 2 mm against about 1.5 mm

Contested 8 GB and 10 GB Nobody has posted a per-location map, so measure your own card before you order. The readings, in the order posted:

@meromgreen (2026-07-30) measured the stock pads at 2 mm with calipers and recommends soft 2 mm pads; on 2026-07-31 he clarified the original pads were all 2 mm, measuring 1.7 to 1.8 mm compressed, after @samuel_hayden. had guessed 1 or 1.5 mm and @meatsus 1.5 mm squeezed. @endsjeremy: many installs need several thicknesses, so check that each pad makes contact without being crushed. @desire13.37 took part in the thread. Source: @desire13.37, 2026-07-30
Source: @endsjeremy, 2026-07-30
Source: @meromgreen, 2026-07-30
Source: @meatsus, 2026-07-31
Source: @samuel_hayden., 2026-07-31
Source: @meromgreen, 2026-07-31
@letterc67 asked the thickness on 2026-08-08; @ju5t5omedude answered with a screenshot whose content is not in the export. Source: @letterc67, 2026-08-08
Source: @ju5t5omedude, 2026-08-08
@turbine (2026-08-12) replaced his with 2 mm pads and @meatsus said someone had measured 2 mm; @bayley_w noted these cards are easier to service than V100s because there is no exposed die. Source: @turbine, 2026-08-12
Source: @bayley_w, 2026-08-12
Source: @meatsus, 2026-08-12
2026-08-13: @sealclubber: 1.5 or 2 mm pads (2 mm squishes to 1.5 mm) and any decent non-conductive paste. @cute_sc2: the stock pads are very soft and compress a lot, so softer pads are better. @rogersmith2026 had to trim a pad for clearance so it could compress, used 2 mm everywhere, saw the core drop less than hoped and planned to try 1.5 mm. @derick2k planned regular paste for now and PTM7950 on the core plus putty elsewhere once he water-cools. @catchyuser asked about Arctic MX-7 and Aairhut 13 W/mK 2 mm pads (unanswered). Source: @rogersmith2026, 2026-08-13
Source: @derick2k, 2026-08-13
Source: @catchyuser, 2026-08-13
Source: @sealclubber, 2026-08-13
Source: @rogersmith2026, 2026-08-13
Source: @cute_sc2, 2026-08-13
@quark2world (2026-08-15) has taken cards apart: all pads are about 2 mm, and pads that read 1.5 mm are the same 2 mm pads compressed. On 2026-08-17: his stock pads were 2 mm while those supplied with his water block were 1.5 mm, and 1.5 mm may still make contact. @catchyuser, @mrmannyatx and @sealclubber took part in those threads. Source: @catchyuser, 2026-08-15
Source: @quark2world, 2026-08-15
Source: @mrmannyatx, 2026-08-17
Source: @sealclubber, 2026-08-17
Source: @quark2world, 2026-08-17
@danske1 (2026-08-19) used 2 mm Gelid pads, which squish a little. @schakaqt, @desire13.37 and @meatsus took part in the thread. Source: @schakaqt, 2026-08-19
Source: @desire13.37, 2026-08-19
Source: @meatsus, 2026-08-19
Source: @danske1, 2026-08-19
@johnnydaboy (2026-09-01) repadded to the original layout with 1.5 mm Arctic TP-3 and separately tried 3 mm TP-3 compressed down (he read that TP-3 compresses safely to about 2 mm). He saw no difference and was still throttling while he waited for new blowers. @meromgreen: the pads should be 2 mm. Source: @johnnydaboy, 2026-09-01
Source: @meromgreen, 2026-09-01
@quark2world (2026-09-07) measured the stock pads on another card at about 1.5 mm. @meromgreen and @meatsus took part in the thread. Source: @quark2world, 2026-09-07
Source: @meromgreen, 2026-09-07
Source: @meatsus, 2026-09-07
@lenxtroz (2026-10-12), measuring from his PCB and die photos: VRAM pads 2 mm, VRM pad size undetermined. @emogop, replying to the PCB photo: MOSFETs 2 mm, coils 1 mm. @con3po7016 and @djpandaman had both heard 2 mm (shore hardness unknown). Source: @lenxtroz, 2026-10-12
Source: @emogop, 2026-10-12
Source: @con3po7016, 2026-10-12
Source: @djpandaman, 2026-10-12
@markshowa (2026-10-18) asked for a measured per-location comparison; none has been posted. Source: @markshowa, 2026-10-18

Pad thickness under water blocks: 1 mm, 1.2 mm or 1.5 mm by the instructions

Contested 8 GB and 10 GB Block makers' instructions disagree with each other and with the stock pads:

@globalapathy (2026-10-11 and 10-12): his water block instructions call for 1.2 mm while the listing says 1.5 mm; he bought 1.5 mm. @gpmickey found mixed information online (1.5 against 2 mm) and ordered 1.5 mm for stock heatsinks, noting some people use 2 mm squishy pads. @civiloid and @cmurrayis took part in the thread. Source: @globalapathy, 2026-10-11
Source: @gpmickey, 2026-10-11
Source: @civiloid, 2026-10-11
Source: @cmurrayis, 2026-10-11
Source: @globalapathy, 2026-10-12
@quark2world (2026-09-07) padded his EKWB A100 block with 1 mm pads per the EK manual; the EK layout pads some spots that were bare on the stock cooler and leaves bare some that were padded. Source: @quark2world, 2026-09-07
@markshowa (2026-10-19) chose T-Global TG-A4500-100-100-1.5 pads (1.5 mm) for the VRM, DrMOS and PMIC after finding 1.5 mm in Bykski's water block instructions, stated as constant across those parts; whether the stock copper cooler uses the same thickness went unanswered. Source: @markshowa, 2026-10-19
@gnasherrr (2026-10-25) added 2.5 mm pads behind the MOSFETs and inductors under a TinyWhale block. Source: @gnasherrr, 2026-10-25

Other thickness claims: 1 mm, 0.5 mm and 3 mm on inductors, without caliper readings

Contested 8 GB and 10 GB Claims nobody backed with a measurement, in the order posted:

In a thread started by @zombielurker (2026-09-20), @endsjeremy had heard the pads are 1 mm (soft, compressible); @prometasfw was fairly sure they are 2 mm (he cut strips from a 100 x 100 mm pad); @kio063218 has seen 1.5 mm and 2 mm for VRM and memory and 3 mm for inductors, and asked for caliper measurements. @danske1: 2 mm soft pads work, avoid hard pads, and stock may not have had pads over the inductors. @quark2world: the original white pads are 2 mm. Source: @zombielurker, 2026-09-20
Source: @endsjeremy, 2026-09-20
Source: @prometasfw, 2026-09-21
Source: @kio063218, 2026-09-21
Source: @danske1, 2026-09-21
Source: @quark2world, 2026-09-21
A comment under "The Repastening" video says 1.5 mm, while @con3po7016 heard they are all 2 mm (2026-10-16); @trustinter took part in the thread. Source: @trustinter, 2026-10-16
Source: @con3po7016, 2026-10-16
@robbbbbb7973 (2026-10-21) did not measure because everyone said 2 mm at the time, and now people say 1.5 mm; @markshowa has also heard 1 mm and 0.5 mm, while a YouTube teardown just replaced them without care. Source: @markshowa, 2026-10-21
Source: @robbbbbb7973, 2026-10-21

The caliper readings above support 2 mm on most cards; nobody posted a measured 1 mm or 3 mm stock pad.

Too-thick pads broke a power line: never stack pads

Warning Pads that were too thick broke a power line on one of @finishtilt14's cards, which needed repair (2026-09-16). @ins_mniac: too much pad can warp the PCB and stress the memory modules, causing GPU errors, while too little leaves the card throttling; never stack pads, and tighten evenly (2026-09-01). @m3gagluk: over the lowest power inductor a 2 + 2 mm stack stressed both the pads and the board; it needs a single pad thinner than the 4 mm gap. After a delid the other pads must come down to 1.5 mm (see the delid warning above).

8 GB and 10 GB
Source: @m3gagluk, 2026-08-15
Source: @ins_mniac, 2026-09-01
Source: @finishtilt14, 2026-09-16

Pad placement: what the heatsink touches, and the inductor gap over 4 mm

@meatsus (2026-07-31): the package has an IHS with thermal paste between the die and the IHS. @samuel_hayden.: the heatsink contacts only the GPU and the HBM on the interposer; the VRMs sit around the core, and pads go on the small black rectangles to the left and right of the parts marked LR22. @meromgreen (2026-08-15) found raised heatsink areas with no pads that an A100 water block manual covers, plus contact points on the top shroud, and added 2 mm pads on all of them except the area to the right of the capacitors over the power inductors, where the gap is over 4 mm. Gelid GP pads do not stick well and fall off when the shroud is removed; @danske1 planned 2 mm pads on the shroud contacts. @quark2world (2026-08-16) photographed the pad positions he needed on the bare board when fitting a water block: pads sit on the tiny (about 0.5 x 0.5 mm) 12 V components around the core, not on the big capacitors, and he suspects some pads also serve as electrical insulation. @skirtchaser took part in the thread.

Reported
Source: @meatsus, 2026-07-31
Source: @samuel_hayden., 2026-07-31
Source: @meromgreen, 2026-08-15
Source: @danske1, 2026-08-15
Source: @quark2world, 2026-08-16
Source: @skirtchaser, 2026-08-16

Were the bare contact points on the stock heatsink meant to be padded?

Contested 8 GB and 10 GB Many cards arrive with heatsink contact points that carry no pad. The views, in the order posted:

Pad them: @quark2world (2026-08-16): the stock cooler has contact points in the same places as a water block's pads but missed about one third of the pads, so when you redo a card, check the small cooler positions that came without pads. Source: @quark2world, 2026-08-16
Leave some bare: @danske1 (2026-09-01) noticed pads added over capacitors where there were none from the factory, which @johnnydaboy removed; a diagram by Alex1997, posted in the thread, marks which heatsink protrusions need no pad. @cressida1881 took part in the thread. Source: @danske1, 2026-09-01
Source: @cressida1881, 2026-09-01
@cressida1881 (2026-09-07) found some contact blocks on the stock heatsink that never had pads or paste, only mining dust, so they are probably not meant to be padded; @robbbbbb7973 confirmed the layout matches his card, and @quark2world posted another card's layout and concluded it is either stock or every card came from the same miner who replaced the pads wrongly, but assumes stock. Source: @cressida1881, 2026-09-07
Source: @robbbbbb7973, 2026-09-07
Source: @quark2world, 2026-09-07
@quark2world (2026-09-21): about 1 mm pads could be added where the cooler should make contact but had none; pads belong on SMD parts, maybe not on the small capacitors around the chips. Source: @quark2world, 2026-09-21

Thermal putty and liquid pads

Putty instead of pads: Upsiren UTP-8 and UTP-X, Laird 607

Putty sidesteps the pad-thickness question. @wakeupmbrc1 recommends putty or liquid pads such as Laird 607 or Upsiren UTP-8 / UTP-X; 100 g of UTP-8 cost him about $25 on Ozon (Taobao is cheaper) and covers 7 to 8 RTX 5090s, so 100 g is plenty for 12 cards, which is what @gpmickey had asked about. @civiloid: Upsiren's official AliExpress shop sells UTP-8, UTP-X and decent PTM7950 clones; UTP-X is rated 16.8 W/mK against 14.8 W/mK for UTP-8 and has been made for about 1.5 years, the two perform about the same in tests, and UTP-X is slightly more compressible. @darcidride planned UTP-8 on the VRAM of all 23 of his cards, for its rated 14.8 W/mK, because it adapts to the 1 mm or 2 mm pad gaps on these cards, and for a claimed life of 5 years or more, with phase-change material or paste on the die (2026-09-22; @_arcanum_ took part). @civiloid used Upsiren putty instead of pads on his stock-cooler reassembly, and @_syn_thetic_ uses it front and back under his water blocks, with Kapton tape around power cables and solder areas, and recommends Duronaut later so the cards never need reopening. @djpandaman was considering putty, and @fiiireface plans PTM7950 on the core and putty on the power parts.

Reported 8 GB and 10 GB
Source: @gpmickey, 2026-09-15
Source: @wakeupmbrc1, 2026-09-15
Source: @civiloid, 2026-09-15
Source: @darcidride, 2026-09-22
Source: @_arcanum_, 2026-09-22
Source: @civiloid, 2026-10-11
Source: @fiiireface, 2026-10-11
Source: @djpandaman, 2026-10-12
Source: @wakeupmbrc1, 2026-10-16
Source: @_syn_thetic_, 2026-10-22
Source: @_syn_thetic_, 2026-10-25

Putty must be non-conductive, and a thick layer insulates

@ins_mniac: any putty must be non-conductive, and thermal coefficients should be compared, since the goal is high conductivity with a minimal layer; too thick a layer insulates, a difference he has seen reach tens of degrees on other water blocks. A pad diagram posted in the same thread marks some spots as "liquid thermal pad (ie. thermal paste)".

Reported 8 GB and 10 GB
Source: @cressida1881, 2026-09-01
Source: @ins_mniac, 2026-09-01

Phase-change material and liquid metal

Is phase-change material worth it on top of the IHS?

Contested 8 GB and 10 GB The views, in the order posted:

@chechu10 (2026-08-12) asked whether PTM7950 beats paste. @quark2world: it is fine but not worth buying specially, because a large IHS sits between the die and the cooler. @skirtchaser agreed the main bottleneck is the paste under the IHS (die to IHS), not the IHS-to-heatsink layer. @sealclubber took part in the thread. Source: @chechu10, 2026-08-12
Source: @quark2world, 2026-08-12
Source: @skirtchaser, 2026-08-12
Source: @sealclubber, 2026-08-12
@quark2world (2026-08-15 and 08-16) found left-to-right grooves you can feel with a finger in the stock cooler's copper plate, unlike a 20-year-old Heatkiller cold plate, and then recommended phase-change material instead of paste. @ju5t5omedude suggested lapping, though he thinks the marks are surface scratches rather than milling marks and may not matter much; @danske1 noticed the copper looks as if it once contacted an exposed die. @dalzyyy linked a card repasted with paste on the memory instead of putty (possibly not a 170HX), and @meatsus planned to test aluminium foil instead of paste on the IHS; no result was posted. Source: @danske1, 2026-08-15
Source: @dalzyyy, 2026-08-15
Source: @meatsus, 2026-08-15
Source: @quark2world, 2026-08-16
Source: @ju5t5omedude, 2026-08-16
@spale777 (2026-08-22) would not run these cards without phase-change material. @_arcanum_: fine if you have some to spare, but 250 W over an IHS that large is not much of a pump-out concern, and unlike old Radeon HBM cards there is no die-versus-HBM height problem, because the 170HX has an IHS. @fff7d1bc took part in the thread. Source: @_arcanum_, 2026-08-22
Source: @spale777, 2026-08-22
Source: @fff7d1bc, 2026-08-22
@ghostinthedatacenter_87281 (2026-10-11) used phase-change material on his 8 GB card and saw a huge temperature decrease; @fiiireface plans PTM7950 on the core, and @gpmickey took part in the thread. Source: @ghostinthedatacenter_87281, 2026-10-11
Source: @fiiireface, 2026-10-11
Source: @gpmickey, 2026-10-11
@skirtchaser (2026-10-18): PTM7950 is pretty much the only interface guaranteed to work on dies with HBM, because the die and IHS are too uneven for liquid metal or regular paste. His position moved from "the paste under the IHS dominates" in August to this one. Source: @skirtchaser, 2026-10-18
@civiloid (2026-10-18) disputes the unevenness argument, since there was paste there before, but agrees a phase-change material is worth it: it is not a thermal pad, and it performs close to liquid metal without the corrosion, short-circuit or gravity-escape problems; Honeywell is the reference and the others are slightly worse. @globalapathy's thermal table places phase change under liquid metal but slightly above graphene sheets. @xeax took part in the thread. Source: @civiloid, 2026-10-18
Source: @globalapathy, 2026-10-18
Source: @xeax, 2026-10-18
@markshowa and @robbbbbb7973 are going with the Laird sheet on the die, and @markshowa plans an A/B test, since the 170tune documentation says a cooler HBM2e gives higher clocks and stability. Source: @markshowa, 2026-10-18
Source: @markshowa, 2026-10-21
Source: @robbbbbb7973, 2026-10-21

Laird A18172-08 sheet: about $36 for ten cards

When @stormlrd. was about to re-pad 20 cards and asked about PTM7950 against the Laird part, @globalapathy ordered the Laird A18172-08 phase-change sheet from DigiKey: 0.2 mm thick, 228.60 x 228.60 mm, about $36, roughly ten cards' worth. By his account it is comparable to Honeywell PTM7950 but cheaper per area, with slightly lower conductance, lasts longer than paste and does not pump out. He also ordered Gelid TP-GP04-C 1.5 mm pads rated 15 W/m-K. The parts arrived on 2026-10-13; no results have been posted.

Reported 8 GB and 10 GB
Source: @globalapathy, 2026-10-09
Source: @stormlrd., 2026-10-09
Source: @globalapathy, 2026-10-11

PTM7950: price, counterfeits and handling

@irythros (2026-08-22): PTM7950 has about 35% higher thermal conductivity; a 40 x 80 mm sheet costs USD 17 (3,200 mm²) and a 100 x 160 mm sheet USD 40 (16,000 mm²), cheaper than the LTT pad at USD 100 for 32,000 mm² (listings: Amazon, Amazon). Keep it in the fridge before applying, make the surfaces very clean, and follow the Thermal Grizzly PhaseSheet application video (YouTube). @skirtchaser finds most AliExpress PTM fake, and @globalapathy warns PTM7950 is heavily counterfeited; LTT sells genuine PTM, and @quark2world points EU buyers to the Thermal Grizzly PhaseSheet (50 x 40 mm). Size: the die area needs a 46 x 46 mm piece and most consumer packs are smaller in one dimension, so @civiloid cuts pieces from 50 x 50 mm sheets and uses Upsiren's PTM clone (PCM-1).

Reported 8 GB and 10 GB
Source: @quark2world, 2026-08-12
Source: @skirtchaser, 2026-08-12
Source: @irythros, 2026-08-22
Source: @globalapathy, 2026-10-11
Source: @civiloid, 2026-10-11
Source: @globalapathy, 2026-10-18

Liquid metal is not recommended

Warning @globalapathy, who has liquid metal on an RTX 5090, warns it is a mess, gets everywhere, needs proper management and still "dries up"; @gpmickey has seen it short components. @globalapathy plans to switch to phase-change material when he services his cards.

8 GB and 10 GB
Source: @globalapathy, 2026-10-11
Source: @gpmickey, 2026-10-11

Package and IHS dimensions

Package 55 x 55 mm, mounting holes 57 x 68 mm, IHS 0.5 mm thick; a 46 x 46 mm IHS unverified

Contested Sizes for cutting phase-change material and for adapter plates, in the order posted:

@gmzombie (2026-07-24) measured the stock heatsink mounting pattern at 57 mm left to right by 68 mm top to bottom, from the centre of each screw, and the die/package at 55 x 55 mm; he suggested a customized RTX 4080 / 4080 Super cooler would make a very strong heatsink. @ju5t5omedude was told by an AI that the closest standard mounts are socket 478, socket 370 or a 60 x 60 pattern (unverified). Source: @gmzombie, 2026-07-24
Source: @ju5t5omedude, 2026-07-24
@dalzyyy (2026-08-30), answering @djpandaman, posted a dimensioned image of the cooler mounting holes, accurate to about 0.1 mm, for adapter plates and custom blocks; @djpandaman separately measured the CMP package at 55 x 55 mm against 55.2 x 55.2 mm for the V100. Source: @dalzyyy, 2026-08-30
Source: @djpandaman, 2026-08-30
@meatsus (2026-09-15): the IHS is 0.5 mm thick. Source: @meatsus, 2026-09-15
@fiiireface (2026-10-06) asked for the size of the central die before ordering PTM7950, without taking his card apart; no answer came in that thread. @globalapathy later said the die area needs a 46 x 46 mm piece and quoted the IHS as 46 x 46 mm "iirc". Nobody has checked that figure against the measurements above. Source: @fiiireface, 2026-10-06
Source: @globalapathy, 2026-10-11
Source: @globalapathy, 2026-10-18

Still unmeasured

Posts with numbers on any of these are welcome on Discord:

A per-location caliper map of fresh, uncompressed stock pads on an 8 GB and a 10 GB card (VRAM, MOSFETs, coils and inductors including the gap over 4 mm, PMIC), with photos.
One before/after repaste on a used card at the same fan duty and power cap, with the card's hours noted.
An A/B of the Laird A18172-08 sheet, PTM7950 and fresh paste on the IHS at matched power and fan speed, core and HBM.
The IHS outline, and whether the top of the IHS sits level with or above the HBM stacks.