Thermal limits

The CMP 170HX runs hot for its power, and it has two temperatures that matter: the GPU core and the HBM memory. Under load the HBM is usually the hotter one, and it is the one that limits memory overclocks and corrupts data first. Keep the core under about 80 C and the HBM under about 80 to 85 C for daily use. The driver itself lists 85 C as the core's maximum operating temperature, 95 C as slowdown, 98 C as shutdown and 95 C as the memory's maximum.

Thresholds

Sensor	Temperature	What happens	Source
Core	about 76 C	throttling starts, according to a rough cooler model; keep core and HBM under 75 C	Theory Source: @zorg33, 2026-10-20
Core	80 C	PixelML's stop rule for core temperature	Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 (Apache-2.0 license)
Core	above about 80 C	GA100 leakage rises with temperature and the card runs away thermally (see thermal runaway)	Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 (no license stated)
Core	85 C	maximum operating temperature (read from the card, 170tune); "the cards will throttle at 85" Source: @matstatman, 2026-10-09; "85 degrees is the maximum, so you're most likely bottlenecked by cooling" Source: @xeax, 2026-10-15; the lab treats 85 C as a wall and notes there is no NVML fan control Source: @darkmaniac7, 2026-10-28 (lab notes)	Source: repo github.com/cachenetics/170tune by cachenetics on GitHub, posted by @rh3dai, 2026-10-04 (MIT license)
Core	95 C	hardware slowdown Source: repo github.com/cachenetics/170tune by cachenetics on GitHub; the card falls off the bus and turns off Source: @analyticeth, 2026-10-07	see cells
Core	98 C	"GPU shutdown temp" in `nvidia-smi -q` on a running card	Source: @matstatman, 2026-08-22
Core	100 C	one card reached it by accident, cut its power as hard as it could, and still works; the owner wrote a remote temperature monitor afterwards (screenshots only, no code)	Source: @sojun80, 2026-10-06
HBM	about 75 to 76 C	a memory overclock (NDIV 70 and up) starts to corrupt data; 170tune qualifies NDIV 70 only with HBM at or below about 76 C	Source: repo github.com/cachenetics/170tune by cachenetics on GitHub
HBM	80 C	"hbm doesn't have issues until 80c or so" Source: @replacedriver47, 2026-10-10; "anything above 80 is too hot" Source: @spectakkular, 2026-10-09	see cells
HBM	85 C	PixelML's stop rule for memory; above 85 C JEDEC halves the refresh interval, so keep stock refresh settings	Source: repo github.com/PixelML/club-170hx by PixelML on GitHub Source: repo github.com/vektorprime/170tune by vektorprime on GitHub, posted by @rh3dai, 2026-10-04 (MIT license; fork of cachenetics/170tune)
HBM	95 C	maximum operating temperature Source: repo github.com/cachenetics/170tune by cachenetics on GitHub; "a few minutes" at 95 C corrupts data in the WPR Source: @bigptoughneigh, 2026-10-09	see cells

What nvidia-smi reports on a running card

@matstatman posted the temperature block of nvidia-smi -q from a running 170HX:

Field	Value
GPU current temp	73 C
GPU max operating temp	85 C
GPU slowdown temp	95 C
GPU shutdown temp	98 C
Memory current temp	81 C
Memory max operating temp	95 C

In that sample the memory ran 8 C hotter than the core. The values themselves are what the driver reports; what the card actually does at each of them is disputed (see Where people disagree).

Contested 8 GB and 10 GB
Source: @matstatman, 2026-08-22

How hot can the HBM run: 70, 75, 85 or 95 C?

No damage threshold has been measured on a 170HX. The answers, in the order they were posted:

Contested 8 GB and 10 GB

@_arcanum_ (from memory): GA100 memory is "positively paranoid" about temperature and starts throttling at about 85 C; @sealclubber noted that other memory may reach about 105 C, with sensors often hidden from Windows. Source: @sealclubber, 2026-07-20
Source: @_arcanum_, 2026-07-20
@giaki3003 worried that 75 C memory might be too hot; @meromgreen: 75 C is merely warm and HBM should be fine until about 95 C, pointing at 110 C hotspots on the 3080; @giaki3003 conceded the 3080 did not use HBM and that he was likely recalling a different warning. Source: @giaki3003, 2026-07-22
Source: @meromgreen, 2026-07-22
Source: @._.patrik, 2026-07-22
@robbbbbb7973: HBM's maximum operating temperature is said to be about 85 C, though the card will not throttle or shut down until much hotter; @quark2world aims for 80 C maximum on air and power limits with nvidia-smi if needed; @skyflyingfish suggested 85 C as the limit. @rogersmith2026 warned never to run under full load without a fan, and @danske1's test card hit 90 C when the tape holding its blower peeled off. Source: @rogersmith2026, 2026-08-08
Source: @danske1, 2026-08-08
Source: @skyflyingfish, 2026-08-08
Source: @quark2world, 2026-08-08
Source: @robbbbbb7973, 2026-08-08
@dalzyyy: under 85 C should be fine, but lower is better because HBM cannot be replaced the way GDDR can. @noskillonlyswag could not keep HBM under 80 C in memtest_vulkan even with a full-speed Arctic P12 Pro and a duct (82 C sustained) and left the default throttle target of about 95 C; @quark2world said data centers aim to keep these cards under 70 C. Source: @noskillonlyswag, 2026-08-10
Source: @legitsplit, 2026-08-10
Source: @_arcanum_, 2026-08-10
Source: @dalzyyy, 2026-08-10
Source: @quark2world, 2026-08-12
@_arcanum_: the card thermal-throttles at 85 C and HBM degrades very rapidly at high temperature, hence a lower throttle point than GDDR; 85 C is still safe, but holding there means you are being throttled. @smit1237, on a single 0.8 A fan, saw the core throttle at exactly 85 C with HBM hovering about 85 C; @zorg33 said a thin 12 cm fan is not enough. Source: @smit1237, 2026-08-26
Source: @meatsus, 2026-08-26
Source: @zorg33, 2026-08-26
Source: @_arcanum_, 2026-08-26
@_arcanum_: HBM2(e) thermally degrades in the mid to low 90s C. Source: @_arcanum_, 2026-09-16
@smit1237 runs his cards at 83 C core and considers that fine; @meatsus: the HBM is then likely 90 C or more, and the large die (826 mm2, not counting the HBM) is vulnerable to long-term heat and to heat cycling of the microbump joints. Source: @smit1237, 2026-09-21
Source: @meatsus, 2026-09-21
@con3po7016: 75 C is the top before you risk memory errors and under 70 C is preferable; he pins his fans so cards never pass 70 C. @bigptoughneigh: at 95 C random bit flips occur, going from about 0.03 % errors (over temperature or a marginal clock) to 20 to 30 % memory read-back failures. @robbbbbb7973: 85 C is the safe maximum and the card throttles after that. @sojun80 called @adusts's sustained 83 C not the end of the world but not ideal, and advised reducing power because HBM and GPU share the package. @segmond triggers his fans at 65 C GPU temperature; @lycorascence says HBM tends to run hotter than the GPU. Source: @con3po7016, 2026-09-27
Source: @segmond, 2026-09-27
Source: @adusts, 2026-09-27
Source: @robbbbbb7973, 2026-09-27
Source: @bigptoughneigh, 2026-09-27
Source: @sojun80, 2026-09-27
Source: @lycorascence, 2026-09-27

The 95 C slowdown can reportedly be disabled

@matstatman: "It seems it is possible to disable the 95 degree throttling". No method, numbers or verification were posted, so this is unverified. With bit flips reported at 95 C HBM (above), disabling a thermal limit risks the card.

Reported
Source: @matstatman, 2026-09-20

Where people disagree

Contested

How hot is too hot: @mountains1605 called @arabus_74064's temperatures on Qwen3.8 Next Flash too hot, while @arabus_74064 thought them normal at full load. Source: @mountains1605, 2026-10-09
Source: @arabus_74064, 2026-10-09 @replacedriver47: the SMs are not the problem, the HBM is, and it is more prone to heat damage Source: @replacedriver47, 2026-10-09; @rrickk420: "you're frying your mem" Source: @rrickk420, 2026-10-09; @bigptoughneigh: memory is not always hotter than the core under load Source: @bigptoughneigh, 2026-10-09; @dottoreargus barely reaches 65 C on the same model Source: @dottoreargus, 2026-10-09.
Does the card protect itself? @impactic: it throttles at 85 C and shuts off at 95 C, after which you have to reboot; he keeps his cards in the 70s and has his system stop work at about 90 C, and @mrmannyatx posted a screenshot of the spec (whether 95 C also applies to the memory is unclear). Source: @impactic, 2026-09-08
Source: @mrmannyatx, 2026-09-08 @_arcanum_: unless you do something deliberately dumb, the card should throttle itself; @segmond: "It doesn't"; @_arcanum_ clarified that he meant throttle, not shutdown. Source: @_arcanum_, 2026-09-16
Source: @segmond, 2026-09-16 @cibernox forgot to plug a new card's fan into the motherboard during a stress test and found it at 100 C: it did throttle, but throttling did not keep it under 100 C (both new cards otherwise passed their stress tests with stable temperatures and no errors), the same outcome as @sojun80's 100 C card in the table. Treat throttling as a last line, not as protection. Source: @cibernox, 2026-09-25
Source: @n3bs, 2026-09-25
@impactic thought the card shuts itself down above 95 C; @prefix10k thought 85 C is already bad. Source: @prefix10k, 2026-10-06
Source: @impactic, 2026-10-06
The throttle point: 76 C by @zorg33's model Source: @zorg33, 2026-10-20 against 85 C by @matstatman Source: @matstatman, 2026-10-09 and the card's own maximum operating value Source: repo github.com/cachenetics/170tune by cachenetics on GitHub.

Without airflow the card dies fast

Proven A card without a fan reaches 95 C quickly even at its 40 W idle, falls off the bus and turns off. Source: @analyticeth, 2026-10-07
Source: @nikhilr1696, 2026-10-07 @programmerboi's passive card stopped showing in nvidia-smi at 85 C while idle Source: @programmerboi, 2026-10-16, and a new 10 GB card "clocked out in under 5 mins without cooling. way worse than V100" Source: @programmerboi, 2026-10-25. The heatsink is designed for high-airflow servers and does not work even at idle without it. Source: @civiloid, 2026-10-16
Source: @globalapathy, 2026-10-25 @globalapathy: a Supermicro fan at 50 % pointed roughly at the card was enough, without a shroud. Source: @globalapathy, 2026-10-25
Warning A failed fan means a dead card or worse; @nikhilr1696 only runs these cards supervised. Source: @nikhilr1696, 2026-10-07 Use a decent fan, not necessarily a blower. Source: @cibernox, 2026-10-07

Older reports: minutes, not hours

The card idles at 30 to 45 W with no fan of its own, so it heats up within minutes. @zapirate: no more than 1 to 2 minutes beyond a POST test; @ju5t5omedude: 10 to 15 minutes at idle, later "under about 10 minutes"; @zefrench6961: 10 to 15 minutes from cold at idle. @phoenixone7's card, idling with only case fans and no blower, climbed to 89 C fairly quickly. @smit1237 ran one idle for 10 minutes without a fan and it survived, while @meatsus had assumed a 5 to 10 W idle and left a card idling for hours in a PC with no airflow. @mfj4's first card, idling passively in open air as an eGPU, sat at 43 to 44 C core and 55 to 56 C HBM at about 35 W with no throttling or errors; his second card idled at 47 W and 72 C, and he stopped idling them without airflow.

Proven 8 GB and 10 GB
Source: @sekkit_75019, 2026-08-01
Source: @ju5t5omedude, 2026-08-01
Source: @mfj4, 2026-08-01
Source: @phoenixone7, 2026-08-04
Source: @zapirate, 2026-08-04
Source: @ju5t5omedude, 2026-08-04
Source: @zefrench6961, 2026-08-13
Source: @tchelobh, 2026-08-13
Source: @ju5t5omedude, 2026-08-13
Source: @segmond, 2026-08-26
Source: @smit1237, 2026-08-26
Source: @meatsus, 2026-08-26

A fan that stops leaves the card cooking

Throttling does not save an unattended card. @danske1's card was cooking after its taped fan fell over, and it did not shut itself down. @spuxx_'s test rig (CMP 100-210 cards) lost motherboard power, and with it the fan headers, while the GPUs stayed powered; by morning the cards were too hot to touch. Keep at least one fan on a supply that stays on, such as a server PSU with a breakout board. Safeguards people run: a cron job that kills workloads as temperatures approach 90 C (@danske1), and a persistent monitor script that shuts the server down above 90 C, written after a blower fell off a card (@syntheticblocks).

Proven 8 GB and 10 GB
Source: @danske1, 2026-08-13
Source: @spuxx_, 2026-09-16
Source: @syntheticblocks, 2026-09-16

At 85 C core the card throttles everything

Warning @meatsus: 85 C core is the CMP 170HX throttle temperature; by then the HBM is probably around 95 C, and the GPU lowers all clocks and voltages to stay below 85 C. @matstatman's screenshot shows the clocks dropping at 85 C during memtest_vulkan with an undersized fan. A card that sits at 85 C is cooling-bound and its memory is near the temperature where bit flips are reported; add airflow or lower the power cap.

Source: @matstatman, 2026-08-14
Source: @meatsus, 2026-08-14

A card that vanishes from nvidia-smi minutes after power-on has usually overheated

Warning 8 GB @raisbecka9680 unlocked an 8 GB card to about 64 GB with cmpunlocker; several minutes later it vanished from nvidia-smi with "unable to determine device: unknown error". No cooling had been fitted: it was a thermal shutdown, and the card came back after a reboot. @schakaqt: the card overheats and shuts off, sometimes it must cool down considerably before a reboot detects it again, and it needs at least one push and one pull 120 mm fan just to idle. @phoenixone7 puts idle draw at about 35 W, @zefrench6961 measured about 15 minutes at idle with no airflow before shutdown, and @quark2world mentioned idle crashes at about 95 C. @irythros: use forced air and possibly repaste, since used cards often have dried paste. Earlier, @gleb.shnshn's card was hot to the touch before any load; @ju5t5omedude: with a fan the cards should be lukewarm at idle, and he has had cards stay busy in diffusion tasks after the program exited, so watch temperature, memory use and load. Fit the cooling before the first power-on, unlock included.

Source: @gleb.shnshn, 2026-08-28
Source: @ju5t5omedude, 2026-08-28
Source: @raisbecka9680, 2026-09-01
Source: @irythros, 2026-09-01
Source: @schakaqt, 2026-09-01
Source: @phoenixone7, 2026-09-01
Source: @quark2world, 2026-09-01
Source: @zefrench6961, 2026-09-01

Cards cooked by a dry run without cooling

Warning A card was reported dead after a quick "dry run" without cooling, and a RAM stick in the same system reportedly died too. @xtj7: with no airflow the passive heatsink saturates quickly even without load, so the card cooked; @lycorascence and @artistosu suspected the PSU. @josephur advised getting the card back from the repairer even if they call it unrepairable. @segmond warned that many cards will be damaged by owners who ignore cooling and then resold (see Card failure rates and longevity). Never power the card without a fan on it, not even for a quick test.

Source: @segmond, 2026-09-01
Source: @josephur, 2026-09-14
Source: @lycorascence, 2026-09-14
Source: @artistosu, 2026-09-14
Source: @xtj7, 2026-09-14

Throttling behaviour

When the cap or the temperature limit bites, the core clock drops and with it memory bandwidth and tokens per second. On a 10 GB card with all SMs unlocked, the power cap tanked memory bandwidth even with the clock locked. Source: @djpandaman, 2026-10-12 See Undervolting and power limits.
Warning A 250 W cap did not bound a fully SM-unlocked 10 GB card. @djpandaman's card at 122 SMs briefly reported about 400 W of board power during benchmarks with the limit at 250 W (he was unsure whether the reading was measured or calculated), and all of it ends up in the same heatsink and VRM. @xeax worries the VRM MOSFETs get overloaded and doubts he could cool that load. On such a card, watch core and HBM temperatures under real load instead of trusting the cap, and undervolt (see Undervolting and power limits#10 GB cards and Power supplies and cabling#PSU sizing). Source: @xeax, 2026-10-23
Source: @djpandaman, 2026-10-23
Prefill is where cards run hottest: @c133's cards spike over their cap during long prefills Source: @c133, 2026-10-09; PixelML's worst-airflow card reached 80 to 85 C core on long-context prefill at 180 W, which is why they lowered the cap Source: repo github.com/PixelML/club-170hx by PixelML on GitHub.
Check whether the card is in hardware power-brake slowdown: nvidia-smi -q | grep -A1 'HW Power Brake Slowdown'. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub
Reported A card cooled by the fan that came in the Alifond package reached 85 C at full fan speed running a 27B Q8 model on llama.cpp; @xeax: 85 C is the maximum, so it is cooling-bound. Source: @programmer_007, 2026-10-15
Source: @xeax, 2026-10-15
Fixes, in order: improve the air path (see Fans and ducts), lower the cap (Source: @xeax, 2026-10-15: sudo nvidia-smi -i 0 -pl 200), then consider water (Water cooling). These chips run 400 W on SXM4 boards; better cooling is what unlocks more of them. Source: @xeax, 2026-10-15

What temperatures people run at

Cooling	Cap	Hottest reading	Source
Water block, full load	250 W	50 C	Source: @leenodie, 2026-10-26
Water block (Bykski N-TESLA-A100-X-V2), 360 mm radiator, fans at minimum	180 W	45 C	Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub
Water, general	not given	50 to 60 C	Source: @endsjeremy, 2026-10-20
Water, 45 C coolant	under load	about 60 C, indefinitely	Source: @quark2world, 2026-09-15
Water, 20 C coolant	under load	about 45 C	Source: @quark2world, 2026-09-15
Air, unlocked 64 GB card	175 W	60 C core, 65 C HBM	Source: @rrickk420, 2026-10-08
One 80 mm fan per pair, no duct	185 W	63 C	Source: @c133, 2026-10-09
S12038-4K push only, four cards, cold garage	200 W	62 C	Source: @alanoo, 2026-10-26
Blower at 50 to 60 %	200 W	60 C	Source: @68x0, 2026-10-15
P12 Pro on a shroud	175 W	65 C	Source: @rrickk420, 2026-10-07
Sealed funnel	150 W	70 C	Source: @globalapathy, 2026-10-07
Air cooled	not given	under 70 C	Source: @cute_sc2, 2026-09-16
Air, fans below 100 %	250 W	under 75 C	Source: @meatsus, 2026-09-16
Blower design at 60 % fan speed	not given	about 75 C	Source: @dalzyyy, 2026-09-16
Air, fans tuned aggressively (loud at night)	180 W	memory 75 C	Source: @voidgiraffe, 2026-09-16
S12038-4K push-pull, 24 C room	300 W	HBM 72 C	Source: @wakeupmbrc1, 2026-10-16
One push and one pull fan at 60 to 70 %	not given	68 C core, 82 C memory	Source: @toffles3179, 2026-08-29
P12 Pro push and pull on 3-card plenums, 10 minutes full power	full	80 C	Source: @reala7096, 2026-10-08
Full-speed Arctic P12 Pro and duct, memtest_vulkan	not given	HBM 82 C sustained	Source: @noskillonlyswag, 2026-08-10
Stock VBIOS, blower at 80 %, 25 C ambient, gpu_burn	250 W	low to mid 80s C	Source: @_arcanum_, 2026-09-16
Single 0.8 A fan	not given	core 85 C (throttling), HBM about 85 C	Source: @smit1237, 2026-08-26
One card, 6-way inference soak for 10 minutes	200 W	HBM 84 to 85 C, GPU 78 C	Source: repo github.com/cachenetics/170tune by cachenetics on GitHub
Four cards, GLM-5.3-Flash TP4	165 W	HBM 82 to 83 C	Source: repo github.com/PixelML/club-170hx by PixelML on GitHub
Eight cards in a server with a fan wall, same cap, by slot	180 W	58 to 68 C on two slots per CPU, 81 to 84 C on the other two	Source: @darkmaniac7, 2026-10-28 (lab notes)
Four idle cards, one 80 mm blower on a printed duct	idle	core 37 to 38 C, memory 41 to 51 C	Source: repo github.com/PixelML/club-170hx by PixelML on GitHub
Air, quiet fans	idle	55 to 60 C	Source: @robbbbbb7973, 2026-09-15
Passive, open air, eGPU (first card)	idle, about 35 W	43 to 44 C core, 55 to 56 C HBM	Source: @mfj4, 2026-08-01
Passive, open air, eGPU (second card)	idle, 47 W	72 C	Source: @mfj4, 2026-08-01
Sandwiched cards with shared intake	under load	34 to 42 C	Source: @moliminous, 2026-10-11

Operating targets: under 80 C, about 75 C on air, under 70 C on water

@quark2world shuts down any workload once a card passes 80 C, normally keeps cards at or below 75 C under load, and aims for under 70 C at full load on water. @robbbbbb7973 keeps both core and HBM at 80 C max and checks both, because the HBM runs about 10 C hotter than the core during inference. @con3po7016 targets 75 C; @voidgiraffe tuned his fans to hold under 75 C because anything cooler was too loud. @_arcanum_: on the stock VBIOS at the full 250 W, 25 C ambient and a blower at 80 %, gpu_burn puts stock cards in the low to mid 80s, so 75 C on air at 250 W is "pretty damn good". The thread's consensus: 75 C is fine on air but abnormally bad under water. @quark2world also warned that mining-farm cards were often packed eight to a case designed for 130 W GPUs and ran much hotter, which affects their remaining life (see Longevity).

Proven 8 GB and 10 GB
Source: @quark2world, 2026-08-11
Source: @robbbbbb7973, 2026-09-15
Source: @quark2world, 2026-09-15
Source: @con3po7016, 2026-09-15
Source: @voidgiraffe, 2026-09-15
Source: @_arcanum_, 2026-09-16
Source: @cute_sc2, 2026-09-16
Source: @meatsus, 2026-09-16
Source: @dalzyyy, 2026-09-16
Source: @voidgiraffe, 2026-09-16

HBM against core: a 2 to 18 C gap, depending on card, paste and airflow

The main temperature nvidia-smi shows is the core. nvidia-smi -q -d TEMPERATURE also shows the memory (HBM) temperature on the 170HX (a 3090, 4090 and 2070 show N/A there), so read both: a card can look fine while its memory is too hot. How big the gap is depends on who you ask:

Contested

@robbbbbb7973: HBM runs about 10 C hotter than the core under load and during inference. Source: @robbbbbb7973, 2026-08-08
Source: @robbbbbb7973, 2026-09-15
@ppc64_ found HBM well under 10 C above the core after repasting dried stock paste, and wonders whether reports of a 10 to 15 C gap come from paste drying around the edges first; @ju5t5omedude: dried paste lets gaps form under thermal cycling. Source: @ppc64_, 2026-08-11
Source: @ju5t5omedude, 2026-08-11
@matstatman's nvidia-smi -q sample showed memory 8 C above the core. Source: @matstatman, 2026-08-22
@meatsus: HBM usually runs at core plus 5 to 10 C; @smit1237 sees HBM never hotter than the chip, with a 2 C delta. Source: @smit1237, 2026-08-26
Source: @meatsus, 2026-08-26
@toffles3179 saw 14 to 15 C (68 C core, 82 C memory) with one push and one pull fan at 60 to 70 %. Source: @toffles3179, 2026-08-29
@ob88594 monitored HBM under GPU stress and found it follows the core within about 10 C, and that keeping it below 70 C is enough; @xtj7 put the worst case under 10 C and usually under 5 C; @sojun80 had challenged the earlier screenshots because they did not show HBM temperatures. Source: @sojun80, 2026-09-15
Source: @ob88594, 2026-09-15
Source: @xtj7, 2026-09-15
@xtj7: HBM usually runs about 5 C above the core and the worst case he saw posted was about 8 C; he keeps the core under 75 C so the HBM stays under 85 C. Logs from another owner, posted by @mlx4, show HBM 2 to 7 C above the core under load. Source: @xtj7, 2026-09-21
Source: @mlx4, 2026-09-21
@segmond normally sees about 6 C, but one card reached 70 C memory at 52 C core (see the fan-control entry below). Source: @segmond, 2026-09-24
The lab: on its airflow-starved hot cards, core and HBM track each other within 3 C; it reads a core far above its own HBM as dried paste. Source: @darkmaniac7, 2026-10-28 (lab notes)

Hotspot: 8 to 10 C over the core is fine, 20 C or more suggests a repaste

@ju5t5omedude sees the hotspot 8 to 10 C above the GPU core temperature, around 85 to 87 C, and asked for a safe hotspot limit. @jiacheng_liu: an 8 to 10 C delta is fine and 85 to 87 C hotspot is fine; a delta of 20 C or more with conventional cooling is where he starts to think about a repaste. Fans can be driven from the hotspot sensor, but that is louder.

Reported
Source: @ju5t5omedude, 2026-08-16
Source: @jiacheng_liu, 2026-08-16

Thermal runaway above about 80 C

The 170th-Street and club-170hx repositories describe the card as a passive datacenter accelerator that reaches 85 C within 1 to 2 minutes without directed airflow and runs away thermally above about 80 C if not cooled: on the GA100 die, higher temperature raises leakage current, which raises the temperature further. Their standard stop thresholds are 80 C core and 85 C memory, with HBM the warmest sensor. The warning reached the Discord when @kellertreppe123 cited "the github repo" and added that it "might be bs"; @quark2world confirmed thermal runaway is real with these cards but said they still have a shutdown mechanism, and would not run them uncooled even at idle. @syntheticblocks quoted the leakage-current explanation from a GitHub write-up. @cressida1881 asked whether the runaway state is permanent or cleared by a reboot; nobody answered.

Proven 8 GB and 10 GB
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub
Source: repo github.com/PixelML/club-170hx by PixelML on GitHub
Source: @kellertreppe123, 2026-09-03
Source: @quark2world, 2026-09-03
Source: @cressida1881, 2026-09-03
Source: @syntheticblocks, 2026-09-16

A 250 W cap needs two to three times the airflow of 200 W

@zorg33 modelled the stock cooler roughly (with Gemini): a 250 W cap needs about 2 to 3x more airflow than 200 W. The heatsink has no vapor chamber and is undersized for 250 W; real server cooling means fans at 8,000 rpm and more. What the heatsink is made of is disputed. Feeding the cards winter air was rejected because of condensation (see Room and inlet temperature).

Contested 8 GB and 10 GB
Source: @zorg33, 2026-10-20

@quark2world examined one: a nickel-plated copper base with an all-copper contact area over the chip, aluminium and copper mixed elsewhere by a scratch test, a triple heat pipe glued to aluminium fins, and a backplate attached to nothing thermally. Source: @quark2world, 2026-08-20
@rh3dai: the stock heatsink only has copper on the core, with aluminium and heat pipes elsewhere. Source: @rh3dai, 2026-10-07
@alanoo: the stock heatsink is full copper. Source: @alanoo, 2026-10-20

Drive the fans from the hottest HBM or core reading, not a fixed curve

@_studebaker_ proposed tying fan control to HBM temperature: nvidia-smi exposes it, the HBM gets much hotter than the core and throttles decode considerably, and a fan curve can be bound to a motherboard fan header. @ob88594 had keyed his fans to core temperature (full load at about 55 dB). @segmond normally sees about 6 C between core and memory, but on one card the memory kept climbing while the core cycled, reaching 70 C memory at 52 C core, so a controller keyed to the core did nothing; he suspects that card needs new thermal material. @phoenixone7 uses a CoolerControl virtual sensor that takes the maximum of both readings; @lycorascence's alternative is to run the fans at maximum. @con3po7016 pins fan control to a thermal target instead of a curve: every card's memory or core at 75 C max regardless of load, which keeps the fans at the minimum needed (on Gigabyte G292-class servers the floor is still about 3,750 rpm); he drives it from memory temperature because it runs hotter and is the more dangerous one. @alanoo regulates on the highest core or HBM reading across his four cards at 200 W (push only with 12038-4K fans, 62 C max in a cold garage); @wakeupmbrc1: adding pull fans helps a lot and lets you run quieter and cooler. The lab's chassis governor does the same on its fan wall, at the cost that the two hot slots cap all eight cards. See Fans and ducts.

Reported 8 GB and 10 GB
Source: @_studebaker_, 2026-08-22
Source: @ob88594, 2026-08-22
Source: @matstatman, 2026-08-22
Source: @segmond, 2026-09-24
Source: @lycorascence, 2026-09-24
Source: @phoenixone7, 2026-09-24
Source: @con3po7016, 2026-10-15
Source: @voidgiraffe, 2026-10-15
Source: @alanoo, 2026-10-26
Source: @wakeupmbrc1, 2026-10-26
Source: @darkmaniac7, 2026-10-28 (lab notes)

Reading the sensors

nvidia-smi -q -d TEMPERATURE Source: @xeax, 2026-10-15, or for logging nvidia-smi --query-gpu=index,temperature.gpu,temperature.memory,power.draw --format=csv Source: repo github.com/PixelML/club-170hx by PixelML on GitHub. Watch temperature.memory: HBM is usually the warmest sensor under load. Source: repo github.com/cachenetics/170tune by cachenetics on GitHub
The card has about 10 temperature sensors, including one per HBM stack. Source: @matstatman, 2026-10-05 matstatman/hbmmon shows per-stack temperature, but it writes to the HBM debug interface by default and needs a custom driver; use its read-only mode on a stock driver. Source: repo github.com/matstatman/hbmmon by matstatman on GitHub, posted by @matstatman, 2026-10-05 (no license file)
The 10 GB card's memory runs a few degrees warmer than the 8 GB card's (two more half stacks active): 41 C against 37 C, both at about 33 W idle, after a repaste; before the repaste the gap was 13 C, which pointed at the thermal interface. The owner used PTM on the 8 GB card and then ran out; what went on the 10 GB card is not stated. Source: @ghostinthedatacenter_87281, 2026-10-13
Source: @colo2442, 2026-10-13
Source: @djpandaman, 2026-10-13
Source: @ghostinthedatacenter_87281, 2026-10-13
Source: @ghostinthedatacenter_87281, 2026-10-13
A server BMC cannot read the temperature of a card on a generic passive riser, so it will not ramp the fans for it. Source: @fafinet, 2026-10-21

Room and inlet temperature

Your intake air sets the floor. The lab's eight-card server raises its power caps only while the inlet is below 24 C and lowers them above 27 C; at about 2 kW into a room, the room became the limit before the fans did. See Chassis builds. Source: @darkmaniac7, 2026-10-27 (lab notes)
Twelve cards on dedicated circuits with an air conditioner pointed at them is still the warmest and loudest room in @gpmickey's house. Source: @gpmickey, 2026-10-19
Warning Do not feed cold winter air or sub-dew-point coolant to the cards: condensation. Source: @alanoo, 2026-10-20
Source: @gnasherrr, 2026-10-26
Paste and pads matter for the last few degrees; see Water cooling for opening the card and Thermal interface: paste, pads and phase-change.

Longevity

Heat and heat cycling are the main suspects for HBM wear, but nothing has been measured on a 170HX yet. Failure counts are on Card failure rates and longevity; cleaning and repasting on Disassembly and cleaning and Thermal interface: paste, pads and phase-change.

Thermal cycling and HBM heat damage

Asked whether improper cooling causes premature death, @cute_sc2 said yes. @_arcanum_ argued that HBM degrades fast once past its safe temperature, so a badly cooled ex-mining card should already be dead unless it ran very few hours; @giaki3003 noted that the cards were released in September and the mining market was "cooked" by November, so most miners bought them when they were already unprofitable, and @quark2world added that some may never have been used. @bayley_w thought letting HBM thermally cycle under poor air cooling is probably not healthy for its bonds. @meatsus wants fan control that holds cards at a constant 60 C rather than swinging between 39 to 40 C and 60 to 65 C, since every expansion cycle stresses the micro-bump solder joints of the roughly 1,000 mm2 package; lower temperature slows diffusion, and lower voltage and power slow electromigration. He guessed about 5 more years of life at a constant temperature. @zapirate's water-cooled card goes from 30 C idle to 50 to 57 C under load at its 300 W limit, and water reacts more slowly than air. @friendsxxii: enough cooling and a low enough power cap should limit the cycling.

Theory
Source: @_arcanum_, 2026-07-27
Source: @cute_sc2, 2026-07-27
Source: @giaki3003, 2026-07-27
Source: @quark2world, 2026-07-27
Source: @thisisashitsystem, 2026-07-27
Source: @bayley_w, 2026-08-09
Source: @meatsus, 2026-08-13
Source: @friendsxxii, 2026-08-13
Source: @zapirate, 2026-08-13

HBM errors reported above about 75 C

@_studebaker_ noticed memory errors start beyond about 75 C HBM; @jimtjames has never gone above 75 C memory temperature and has seen no errors, and keeps the HBM as cool as possible because it is temperature sensitive. At idle the HBM reads well above the core: @noskillonlyswag saw 51 C HBM at 38 C core, so a core at 65 C likely means HBM above 75 C; @toffles3179 sees HBM about 15 C above the core; @jimtjames: the gap may shrink under load because the HBM does not downclock at idle. @sealclubber suggested a repaste for a large gap. @beamreaching1 read 49 C memory at 30 C core at idle (two 40x40x56 mm fans at about 70 %, about 25 C ambient, idle power never above 31 W), and later 36 and 39 C; @chonsworth's card idles below about 36 C with an open case and a floor fan. @cressida1881 saw 36 C core and 51 C HBM at idle on a card that was never repasted. @jjb01957: core and HBM are almost never the same temperature, and the numbers depend on room, case, fan, shroud, BIOS and wattage; what matters is a stable, non-rising temperature under the real workload, and he aims for the 70s under load. One of his cards runs notably hotter HBM and may need a repaste. Watch both sensors with watch -n 1 nvidia-smi -q -d TEMPERATURE or nvidia-smi --query-gpu=temperature.gpu,temperature.memory --format=csv. See Memory errors and ECC.

Reported
Source: @_studebaker_, 2026-08-05
Source: @jimtjames, 2026-08-05
Source: @noskillonlyswag, 2026-08-05
Source: @sealclubber, 2026-08-05
Source: @toffles3179, 2026-08-05
Source: @beamreaching1, 2026-08-30
Source: @chonsworth, 2026-08-30
Source: @cressida1881, 2026-09-20
Source: @jjb01957, 2026-09-20

One card much hotter than its twin: dirt, old paste or a dead fan

@mrmannyatx had two cards on the same task and airflow: one stayed under 50 C, the other sat around 73 C. @ju5t5omedude advised comparing wattage and fan rpm and suspected degraded paste and pads or lint; the hot card was the dirty one ("Dirty=hotter=likeliest to have degraded paste and pads"). @fallengod9962 saw one card at 92 C under a blower while the other ran cool; @scallion_9883 suggested a repaste, repad and blowout, @ju5t5omedude asked whether both fans were spinning, and the cause was a dead fan. @ju5t5omedude also notes the hotspot reads about 10 C above the GPU temperature. Check the fans first, then clean and repaste.

Proven
Source: @mrmannyatx, 2026-08-10
Source: @ju5t5omedude, 2026-08-10
Source: @fallengod9962, 2026-08-12
Source: @scallion_9883, 2026-08-12

Looping garbage output traced to heat

@catchyuser's Qwen 3.8 27B FP8 model on vLLM got stuck after two or three prompts and produced garbage, then an infinite loop under a second harness. @c133 compared launch settings and asked first whether the card had passed cuda_memtest or a Vulkan memtest; @matstatman noted that the sampling temperature is set per request, not at server startup. The card had passed memtest, and the cause seemed to be heat: it runs above 80 C, up to 85 C, and the server fans are too loud to run faster, so the owner is considering water cooling. When output degrades, rule out both memory errors (memtest, see Testing and burn-in) and heat.

Reported
Source: @catchyuser, 2026-09-02
Source: @c133, 2026-09-02
Source: @matstatman, 2026-09-02
Source: @catchyuser, 2026-09-02
Source: @c133, 2026-09-02

Wear gloves when opening ex-mining cards

Warning @robbbbbb7973 cut his finger cleaning a card and the dirt infected it. @_arcanum_, who has opened many cards, says gloves are a must, especially when getting the board out of the backplate, where the retaining wings on the EPS connector are a badly designed part of the job. @zapirate and @impactic urged seeing a doctor about an infected cut. Wear gloves; see Disassembly and cleaning.

Source: @robbbbbb7973, 2026-08-07
Source: @_arcanum_, 2026-08-07
Source: @zapirate, 2026-08-07
Source: @impactic, 2026-08-07

Dust on ex-mining cards: mould and possibly asbestos

Warning @machinezer0 fell ill after blowing dust out of cards bought from a US seller, while nobody else in the house got sick, and suspected mould spores ("Buyer beware"). @humandrone8721_60165 keeps a turbo blower and FFP2 masks for cleaning. @djpandaman found brown dust on two of his cards from China; @spectakkular warned that dust on cards bought from China may contain asbestos, saying asbestos was only banned in new Chinese warehouses in 2011. Nothing has been tested. Clean with care, wear a mask and avoid breathing the dust.

Source: @machinezer0, 2026-08-29
Source: @humandrone8721_60165, 2026-08-29
Source: @meatsus, 2026-08-30
Source: @djpandaman, 2026-09-24
Source: @spectakkular, 2026-09-24