Cooling and power

The CMP 170HX is a passive data-center card. It has a large copper heatsink and no fan. It was built for a server that forces air through its fins, so on a desk or in a normal case it overheats unless you add that airflow yourself or fit a water block. This page is the short version. The detail lives on the pages listed at the end.

Before you power a card on

Warning Never run the card without forced airflow, not even at idle. An idle card draws about 40 W and with no fan it reaches 95 C quickly; at 95 C it falls off the bus and turns off. Source: @analyticeth, 2026-10-07
Source: @nikhilr1696, 2026-10-07
Source: @nikhilr1696, 2026-10-07 One card on passive cooling stopped showing in nvidia-smi at 85 C while idle Source: @programmerboi, 2026-10-16; another clocked out in under 5 minutes with no cooling. Source: @programmerboi, 2026-10-25 See Thermal limits.
- Proven It happens within minutes. The card idles at 30 to 45 W. @zapirate allowed no more than 1 to 2 minutes beyond a POST test; @ju5t5omedude put idle without a fan at 10 to 15 minutes (later "under about 10 minutes"), and @zefrench6961 also said 10 to 15 minutes from cold. @phoenixone7's card, idling on case fans with no blower, climbed to 89 C fairly quickly. @mfj4's first card idled in open air at 43 to 44 C core and 55 to 56 C HBM at about 35 W, but his second idled at 47 W and 72 C, so he stopped idling cards without airflow. @smit1237 ran one card idle for 10 minutes without a fan and it survived. 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
- Warning "Host off" counts too. @spuxx_'s test rig (CMP 100-210 cards) lost motherboard power, and with it the fan headers, while the GPU supply stayed on; 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. @danske1's card was cooking after its fan fell over, and it did not shut itself down. Safeguards in use: @danske1 runs a cron job that kills workloads as temperatures approach 90 C, and @syntheticblocks wrote a monitor script that shuts the server down above 90 C after a blower fell off his card. Source: @danske1, 2026-08-13
  Source: @spuxx_, 2026-09-16
  Source: @syntheticblocks, 2026-09-16
The power input is one EPS12V socket (the CPU 8-pin, 4x 12 V and 4x ground), not a PCIe 8-pin. Feed it with one EPS cable, or with both inputs of the bundled 2x PCIe 8-pin adapter on two separate PSU cables. Never use pigtails or one cable for both inputs. See Power supplies and cabling and Power and cabling warnings.
Watch the memory (HBM) temperature, not only the core. HBM is usually the hottest sensor under load. (power-cap notes) Source: repo github.com/cachenetics/170tune by cachenetics on GitHub, posted by @rh3dai, 2026-10-04 (MIT license)

What airflow it needs

The fins are dense and the heatsink is long, so the fan has to push air through a restriction. What matters is static pressure and a sealed path, not open-air flow.

Air has to go through the fins, from the end of the card out through the slot bracket. Gaps where air can leak around the card cost you most of the effect. Taping over gridded slot brackets in a 2U server took 10 C off one card at 250 W. Source: @mlx4, 2026-10-20
Push-pull (a fan on each end of the card) helps a lot and lets you run fans slower. Source: @wakeupmbrc1, 2026-10-16
Airflow needed rises steeply with power. A rough model by @zorg33 put a 250 W cap at 2 to 3 times the airflow of 200 W. Theory Source: @zorg33, 2026-10-20
Proven A shroud or duct is mandatory. Case fans, or fans pointed at the card, do not cool it: without a sealed path the air takes the way of least resistance around the heatsink, and consumer cases have too many holes, unlike sheet-metal server cases that are open only at front and back (@quark2world). Generic shrouds will not fit, but CMP-specific ones are sold. @mountains1605 confirmed the stock front fan of an HP Z4 G4 will not cool the card, and @quark2world's case-fan setup was barely enough for about 200 W with every fan at full speed. @mrmannyatx's second card kept climbing at idle with an 80 mm push fan and a 120 mm pull fan until he maxed the fans (43 C at 43 W). Motherboard fan control cannot read the card's temperature from nvidia-smi; @ju5t5omedude uses CoolerControl for fan curves. Source: @mrmannyatx, 2026-08-10
Source: @_arcanum_, 2026-08-10
Source: @ju5t5omedude, 2026-08-10
Source: @bayley_w, 2026-08-10
Source: @c.03937, 2026-08-10
Source: @quark2world, 2026-08-14
Source: @danske1, 2026-08-15
Source: @mountains1605, 2026-08-15
Source: @spale777, 2026-08-22
Source: @quark2world, 2026-08-22
Proven Static pressure, not airflow: ordinary fans fail on this heatsink. The fan has to push air through about 6 inches of dense copper fins (@turbine). An Arctic P12 Pro will not push enough air (@meatsus); @letterc67 cools only about 150 W at 75 C with an Arctic P12 Max. @fallengod9962 cooled two cards with one 1,400 rpm 120 mm case fan and a printed duct, and the second card reached 95 C at idle and crashed. @cressida1881's generic 120 mm fan taped to one card let HBM pass 76 C in memtest, and @_datacat_'s 3,000 rpm axial was not enough where a blower card ran cooler. @vcore2191: axials under 2,000 rpm cannot make the pressure the heatsink needs, and larger fans at the same rpm barely help. An AIO radiator next to the card needs a blower regardless, and the top card will bake (@ju5t5omedude, @_studebaker_). Source: @meatsus, 2026-08-02
Source: @replacedriver47, 2026-08-04
Source: @meatsus, 2026-08-04
Source: @letterc67, 2026-08-04
Source: @fallengod9962, 2026-08-12
Source: @turbine, 2026-08-12
Source: @vcore2191, 2026-08-23
Source: @_datacat_, 2026-09-02
Source: @ju5t5omedude, 2026-09-02
Source: @_studebaker_, 2026-09-02
Source: @polskiekaktus, 2026-09-02
Source: @cressida1881, 2026-09-05
- The target: a server GPU cooling 300 W sees about 10 mmH2O of pressure drop at its required airflow, so aim for a fan or fan stack with about 20 mmH2O static pressure at zero flow. One 120 mm fan that strong is overkill on airflow; use a slow 120 mm followed by a fast 92 x 38 mm, or push-pull. Source: @petrikrohn_26497, 2026-08-12
- What works: 38 mm thick 8,000 rpm Arctic fans are the quietest shown to cool the heatsink (@bayley_w). @meromgreen compared the Arctic S8038-10K (about 50 mmH2O, a loud "screamer") with the S12038-8K (about 37 mmH2O) and suggested push-pull at 50 to 60 % for quieter high airflow; @irythros pointed to the S12038-8K's published P-Q curve (variable down to about 400 rpm). @turbine cools four cards with two Arctic S12038-4K. @mrmannyatx ordered the Arctic S8038-7K, which @spale777 called the best readily available 8038 for the price; @meatsus runs an Arctic S8038 at 60 % with a tape-and-paper duct. @c133's single ducted 80 x 80 x 25 mm fan held 29 C at idle and never passed 65 C over half an hour of Qwen3.8-27B at full power, but is very loud at full speed; @beamreaching1's two 40 mm high-rpm fans per card hold HBM at or below 67 C under synthetic load, also very loud (@intervitens runs two 40 mm fans with hearing protection). Industrial fans and 120 mm blowers were suggested by @artistosu and @tacticalsquid. Source: @artistosu, 2026-08-04
  Source: @tacticalsquid, 2026-08-04
  Source: @meromgreen, 2026-08-04
  Source: @turbine, 2026-08-12
  Source: @bayley_w, 2026-08-14
  Source: @intervitens, 2026-08-14
  Source: @c133, 2026-08-14
  Source: @meatsus, 2026-08-14
  Source: @irythros, 2026-08-22
  Source: @mrmannyatx, 2026-09-02
  Source: @spale777, 2026-09-02
  Source: @beamreaching1, 2026-09-05
- Blowers and push-pull: to hold 250 W on one card the fan must draw more than 8 W; avoid 8025 axials, prefer a 9733-type blower, and a 12032 blower is best (@ob88594, whose final pick was a BUC1412VE00). He keeps fans under about 3,000 rpm for acceptable noise and says a good blower can handle 250 W at 50 dB in a 25 C room. Push-pull roughly halves the pressure each fan must make, so 120 mm axials start to be enough at about 3,000 rpm in push-pull, though blowers are the best option short of water or a new heatsink (@vcore2191). A 120 mm fan pulling from the back of the card works better than one pushing (@wakeupmbrc1). Cheap Arctic 4,000 rpm 120 mm fans at 2,000 to 3,000 rpm are about as loud as a room fan (@mountains1605). See Blowers and Fans and ducts. Source: @ob88594, 2026-08-22
  Source: @stcord, 2026-08-22
  Source: @vcore2191, 2026-08-23
  Source: @ju5t5omedude, 2026-08-23
  Source: @wakeupmbrc1, 2026-09-02
  Source: @mountains1605, 2026-09-05
Reported Airflow figures: about 20 to 25 CFM per card, and about 60 CFM for a pair (@squishysquare, who designs and sells printed shrouds for passive cards). His two cards in a dual shroud with one 56 CFM fan stay at 60 to 65 C under LLM loads, and he called @ikyo's dual-80 mm two-card shroud with two 10,000 rpm fans more than enough. @deejay9983 countered that larger fans move more air more quietly, and @meromgreen stressed static pressure over CFM. Source: @squishysquare, 2026-08-12
Source: @deejay9983, 2026-08-12
Source: @meromgreen, 2026-08-12
Source: @squishysquare, 2026-10-03
Source: @ikyo, 2026-10-03
The room matters too. The lab's eight-card server put about 2 kW into the room; raising its fan wall from 10 % to 25 % did not cool the hottest cards (they stayed at 84 to 85 C) while the inlet air rose from 24 C to 26 C. Reported Source: @darkmaniac7, 2026-10-27 (lab notes). See Chassis builds.
- @_studebaker_ vents his network closet to the attic with an inline duct fan pulling from a crawl space, and limits fresh air near a bay to keep salt out of the heatsinks. Source: @_studebaker_, 2026-08-22
Proven Four ways to cool the card, as @lunaroseoc listed them in July 2026: (1) a printed bracket with a 40 mm fan; (2) a printed single- or double-card shroud with an 80 to 120 mm fan; (3) water cooling, since A100 water blocks fit; (4) the stock shroud removed and a custom cooler fitted, such as a Noctua Threadripper cooler held by gravity or zip ties, because no mounting pressure is needed. @farawaythrower added that the power delivery still needs airflow with option 4. @cute_sc2 runs cards in a Gigabyte G292-Z20 server with very loud 80 mm fans and water-cools the one in his workstation; @xnupanic76's two-fan shroud keeps his card under 60 C with the fans below half power. The table below puts numbers on each. Source: @._.patrik, 2026-07-20
Source: @lunaroseoc, 2026-07-20
Source: @giaki3003, 2026-07-20
Source: @farawaythrower, 2026-07-20
Source: @cute_sc2, 2026-07-20
Source: @xnupanic76, 2026-07-21

Cooling options, cheapest to best

Option	Rough cost	Noise	What people measured	More
One 120 mm or 80 mm fan per pair of cards, taped or clipped to the card ends, no duct	one fan per two cards	low to medium	under 63 C at a 185 W cap with Supermicro FAN-0074L4 80 mm fans Source: @c133, 2026-10-09	Fans and ducts
3D-printed intake plenum with 120 mm server fans (Arctic S12038-4K), push or push-pull	filament plus two fans per four cards	medium	300 W per card at 3,500 rpm with HBM at 72 C in a 24 C room, push-pull Source: @wakeupmbrc1, 2026-10-16	Fans and ducts
Printed housing with a blower per card (9733 or 8025 blower)	one blower per card	loud at full speed	60 C at a 200 W cap with the blower at 50 to 60 % Source: @68x0, 2026-10-15	Fans and ducts
Server chassis with its own fan wall	a used GPU server	loud; the BMC fan curve follows GPU temperature	see Chassis builds	Chassis builds
Dual-card AIO water cooler (two blocks, one radiator)	about $250 to $400 per two cards before shipping (October 2026) Source: @_arcanum_, 2026-10-04 Source: @artistosu, 2026-10-04	quiet	4 cards "cool and quiet" Source: @civiloid, 2026-10-16	Water cooling
Full-cover water block on a custom loop	about $140 to $220 per block plus the loop Source: @peter643, 2026-10-20 Source: @_syn_thetic_, 2026-10-25	quiet	50 C max at 250 W full load Source: @leenodie, 2026-10-26	Water cooling

Power facts

Stock power limit is 250 W on the common 8 GB VBIOS (92.00.67.00.01); some 8 GB cards ship with a 300 W VBIOS (92.00.6D.00.0A). Both share the board part 900-11001-0108-000. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 (Apache-2.0 license)
Most people serve at a cap between 140 W and 200 W. A four-card GLM-5.3-Flash TP4 sweep found 140 W only about 1 % slower than 150 W for 8 % less power Source: repo github.com/PixelML/club-170hx by PixelML on GitHub; 175 W is the knee in the 170tune single-card curve Source: repo github.com/cachenetics/170tune by cachenetics on GitHub. See Undervolting and power limits.
Short spikes go above the cap: up to 380 W on cards capped at 250 W. Source: @n2dcrypt, 2026-10-05 Size PSUs for that.
Idle draw is 30 to 40 W per card stock; patched idle settings reach 23 to 26 W. See P-states and idle power.
Undervolting (shifting the V/F curve) cut power 24 % at a fixed 1,350 MHz on the 170tune reference card. It needs a VBIOS that exposes the VF offset; the 250 W 8 GB VBIOS and the 10 GB VBIOS do not. Source: repo github.com/cachenetics/170tune by cachenetics on GitHub

Warnings

Things that melted, caught fire or killed a card. More on Power and cabling warnings.

Power and wiring

Warning On a multi-GPU board, plug in every auxiliary power connector the board has. Without the extra slot power, the board's two 12 V wires that feed the PCIe slots can overload and melt, which has caused house fires (@skirtchaser). With 4 or 5 GPUs on a basic board such as a Z790, use risers with their own 6- or 8-pin power, or the slot will burn; @lunaroseoc's W790E-SAGE refused to run more than one PCIe card without its supplementary power plugged in. She runs one PSU for the GPUs and one for CPU, board and add-in cards, ideally on a board with dual 24-pin or dual-PSU support. @globalapathy: skipping the extra PCIe power plug can make multi-GPU systems unstable. Source: @lunaroseoc, 2026-07-19
Source: @simba6959, 2026-07-19
Source: @skirtchaser, 2026-07-19
Source: @globalapathy, 2026-10-02
Warning SATA power is unsafe at GPU loads. In @ezybr0's tests SATA-powered OCuLink slot adapters were safe for only about 25 W through the slot (he had not tested a CMP); SATA is not made to deliver 75 W, so prefer a 6-pin powered adapter (@ju5t5omedude, @segmond). SATA connectors with plastic molded around the wires can start house fires: the plastic heats, melts and begins conducting. If you must use SATA adapters, use crimped connectors with a clear barrier between wires; Molex does not have this failure, and server PSU breakout boards are fine (@skirtchaser). Source: @ezybr0, 2026-08-25
Source: @ju5t5omedude, 2026-08-25
Source: @segmond, 2026-08-25
Source: @skirtchaser, 2026-09-03
Source: @con3po7016, 2026-09-03
Source: @segmond, 2026-09-03
Warning Cheap power adapters and cables melt under sustained load. @smit1237's cheap adapters on MI50s melted under 200 W sustained training loads; he now makes his own cables. @xeax had a cable start burning at the GPU end on a 4-card 1,400 W rig and caught it in time. Continuous load is much harder on connectors than load that ramps up and down, which is how CMP cards run 24/7 (@_studebaker_). Source: @smit1237, 2026-08-20
Source: @_studebaker_, 2026-08-20
Source: @xeax, 2026-08-20
Warning Keep open power wiring away from anything flammable. @giaki3003 runs open-wire risers on server-PSU-to-PCIe power converters on cardboard boxes and said the cardboard caught fire five times (half-joking, but the risk is real). Source: @giaki3003, 2026-08-03
Warning Mining boards with an onboard 12 V PDU and cheap mining PSUs are a fire risk. @smit1237 saw a miner feed four RTX 3090s through a single PCIe 12 V cable into such a board, which melted the cable from the PSU; cheap mining PSUs may be rated 2 kW while using a transformer from a 600 W unit. @giaki3003 said every cheap mining PSU or board he bought from China ended in a fire, and @1a2b3c0967 agreed. Source: @giaki3003, 2026-08-23
Source: @smit1237, 2026-08-23
Source: @1a2b3c0967, 2026-08-23
Warning Do not feed one card from two unrelated PSUs. @mfj4 powered a card's EPS input from his eGPU supply while the motherboard fed the slot, and it worked, but switching supplies on one device can backfeed each other's 12 V rails; only systems designed for several PSUs have hardware to prevent it (@ju5t5omedude, @bayley_w). For a server PSU feeding a pico PSU on the same board, @endsjeremy warned that paralleled supplies without a shared bus bar push current into each other, and suggested a redundant breakout board or a board made for two CRPS inputs; @quark2world mentioned sync boards or dual-PSU starter cables. Nobody measured the backfeed. Source: @mfj4, 2026-08-25
Source: @ju5t5omedude, 2026-08-25
Source: @bayley_w, 2026-08-25
Source: @ju5t5omedude, 2026-09-03
Source: @endsjeremy, 2026-09-03
Source: @quark2world, 2026-09-03
Source: @segmond, 2026-09-03

Fans and fan headers

Warning Current ratings on cheap AliExpress blowers are often fake. @skirtchaser bought several BFB0712HF (1.8 A) fans and all were fakes to varying degrees: real draw was 0.4 to 0.7 A, peaking around 1.2 A with airflow almost fully blocked, and even fans from a brand's official store turned out fake. @dalzyyy's blower is labelled 3 A, which he and @ju5t5omedude doubted. @squishysquare: cheap AliExpress fans overstate their current, real full speed being about 1 to 1.2 A. @ju5t5omedude noted that fans sold under one 75 mm part number ranged from 72 to 76 mm. Do not size a fan by its label. Source: @ju5t5omedude, 2026-08-04
Source: @dalzyyy, 2026-08-04
Source: @skirtchaser, 2026-08-04
Source: @a.civardagezen, 2026-08-04
Source: @squishysquare, 2026-08-21
Warning High-current blowers can burn motherboard fan headers and the card's own fan output. @fallengod9962 ran blowers at maximum from the BIOS; one killed the motherboard header, the fan stopped silently and the card reached 92 C. These blowers pull 1.85 A or more if allowed; one tripped over-current on a three-channel controller, so @ju5t5omedude caps his curve at 66 %, and a header stuck in over-current protection may need a hard reset. 97 x 33 blowers can pull up to 2.9 A (@vcore2191). Check the header rating, or power the fan from a hub or the PSU. For a fan header soldered onto the card itself, stay under 1 to 1.5 A. @quark2world recommends a script that controls fan speed and stops workloads at around 80 C. See Fans and ducts. Source: @fallengod9962, 2026-08-12
Source: @quark2world, 2026-08-12
Source: @ju5t5omedude, 2026-08-12
Source: @ju5t5omedude, 2026-08-14
Source: @_studebaker_, 2026-08-22
Source: @ju5t5omedude, 2026-08-23
Source: @vcore2191, 2026-08-23

Opening the card

Warning Removing a stuck stock cooler: twist, never pry. The old paste can hold like glue, and capacitors and inductors next to the package break off: @meatsus snapped electrolytic capacitors off one card, and @_arcanum_ knocked an inductor off another and soldered it back. Hold the card with the PCB over your palm and your fingers on the sides of the cooler, then twist back and forth in the plane of the die. Never pull or pry the halves apart: there is no gap for a tool. Check that no screws were missed; contact cleaner spray, or 99 % IPA squirted into the gap with a needle, can help loosen it. See Disassembly and cleaning. Source: @periloustech, 2026-08-05
Source: @coolgod_72592, 2026-08-09
Source: @ju5t5omedude, 2026-08-09
Source: @sealclubber, 2026-08-09
Source: @diabeetee, 2026-08-09
Source: @meatsus, 2026-09-07
Source: @_arcanum_, 2026-09-07
Source: @ins_mniac, 2026-09-07
Source: @endsjeremy, 2026-09-07
Source: @finishtilt14, 2026-09-16
Source: @marklintonca, 2026-09-16
Source: @meatsus, 2026-09-16
Source: @neapopa, 2026-09-16
Source: @_arcanum_, 2026-09-24
Source: @djpandaman, 2026-09-24
Warning Delidding (removing the heat spreader) is risky and there are no ready-made tools. @meatsus chipped the edge of the package substrate on his first attempt (the card still works); every early delid was an accident while removing the cooler (@_arcanum_); old delid guides for other NVIDIA chips had a failure rate (@skirtchaser). The card has little edge space and many capacitors close to the heat spreader (@xanthefin). @quark2world: it can damage the chip and is only needed to fit A100 80 GB-style blocks. Source: @meatsus, 2026-08-01
Source: @danske1, 2026-08-01
Source: @artistosu, 2026-08-01
Source: @meatsus, 2026-08-03
Source: @quark2world, 2026-08-03
Source: @_studebaker_, 2026-08-21
Source: @ju5t5omedude, 2026-08-21
Source: @cute_sc2, 2026-08-29
Source: @skirtchaser, 2026-08-29
Source: @_arcanum_, 2026-08-29
Source: @xanthefin, 2026-08-29
- Warning The first deliberate delid in the cooling channel (@meatsus, 2026-09-15): cut only along the sides of the heat spreader, not the top and bottom. The adhesive is soft silicone, and heat does not help (a piece heated to 200 C for 2 minutes did not change). If the heat spreader goes back on, glue it or print a bracket for it: one that slides sideways kills the small parts around the package. @ins_mniac rates delidding a grave risk for a marginal benefit, because the die is large and prone to cracking. A longer write-up by @meatsus exists (pointer, waiting for the author); see Discord. Source: @meatsus, 2026-09-15
  Source: @dalzyyy, 2026-09-15
  Source: @ins_mniac, 2026-09-15
Warning A swapped heatsink must cool the VRM too. @gmzombie test-fitted an RTX 4080 heatsink ("fits kinda"); @zapirate warned that the VRM needs cooling as well as the core, especially because the 170HX VRM is stripped down compared with the A100's. The same applies to option 4 in the list above. Source: @gmzombie, 2026-07-21
Source: @zapirate, 2026-07-21
Source: @quark2world, 2026-07-21
Source: @_arcanum_, 2026-07-21

Pages in this section

Fans and ducts: fan models, measured temperatures, printable shrouds and ducts, fan control.
Blowers: 97 mm, 80 mm and 120 mm blowers, sizing, adapters and noise.
Printable shrouds and ducts: printable housings, shrouds and ducts.
Water cooling: blocks that fit, AIOs, installation, pads and paste.
Thermal interface: paste, pads and phase-change: repasting and pads.
Disassembly and cleaning: opening the card, freeing the cooler, cleaning.
Undervolting and power limits: V/F offset, power caps, measured watts against tok/s.
P-states and idle power: idle draw and how to lower it.
Power supplies and cabling: the EPS input, cable rules, PSU sizing per card count.
Thermal limits: core and HBM thresholds, throttling, what people run at.
Chassis builds: multi-card builds, fan walls and power distribution.