Blowers

A blower (a centrifugal or "turbine" fan) fixed to the intake end of the card is the most common way to air-cool a CMP 170HX outside a server chassis. It builds pressure at a lower speed than an axial fan, keeps the card two slots wide and does not depend on case airflow. The costs are length (a card with a blower adapter can pass 40 cm) and a sound that some owners cannot live with. The usual choice is a 97 x 97 x 33 mm (9733) 12 V 4-pin PWM blower rated 1 A or more, on a printed adapter or taped on, with its speed driven from the card temperature. Axial fans, shrouds, fan control and header current limits are on Fans and ducts; printable files are indexed on Printable shrouds and ducts.

In short

1. Buy a 9733 12 V 4-pin PWM blower. San Ace B97 clones work and cost $5 to $12; the brand does not matter, but cheap listings overstate their current (see the warning below).
2. A fan that really draws more than about 7 to 8 W holds one card at 250 W. A 0.47 A Nidec G10D is the bare minimum, and a 1 A 9733 is enough for one card up to 250 W.
3. Fix it with a printed A100-style adapter, or tape it on with aluminium HVAC tape or Kapton tape, so the air has to go through the fins.
4. Drive it from the card temperature (a USB PWM controller with CoolerControl, or motherboard headers) and cap the speed: most posted setups run between 40 and 70 %. Check the header current limits on Fans and ducts before you plug a high-current blower into a motherboard.
5. Expect about 60 to 80 C at 175 to 250 W, depending on the fan and how well the adapter seals. If the noise is too much, cap the card at 175 to 180 W (Undervolting and power limits).

Blower or axial fan?

Contested 8 GB and 10 GB

• @ju5t5omedude: stacked two-slot cards leave no gap for a fan to draw intake between them, so he favours a printed housing with integral blowers, each tied to its card's temperature. Source: @ju5t5omedude, 2026-07-24
• @a.civardagezen: a single larger 120 mm fan ducted across the fins is simpler, quieter and more robust, and these server cards omit blowers for reliability reasons. Both setups reportedly cool adequately. Source: @a.civardagezen, 2026-07-24
• @dalzyyy suspects the sparse fins do not need much static pressure, so the real difference is airflow (CFM). Source: @dalzyyy, 2026-08-10
• @noskillonlyswag asked whether a blower beats a 3,000 rpm Arctic P12 Pro with a duct. @farawaythrower: blowers are prized for static pressure. @ju5t5omedude: absolutely; P12s were always barely adequate, airflow is what matters, and fan power draw is the quickest proxy for it. Source: @noskillonlyswag, 2026-08-10
  Source: @farawaythrower, 2026-08-10
  Source: @ju5t5omedude, 2026-08-10
• @vuhuycan0322 asked why not use 120 x 120 x 38 mm 6,000 rpm server fans; the answer was noise. Source: @vuhuycan0322, 2026-08-10
• @endsjeremy posted a chart: a properly sized centrifugal blower reaches enough static pressure earlier than an axial fan, so it can be quieter. @ju5t5omedude: blowers can be louder, but their pitch is less annoying. Source: @endsjeremy, 2026-08-10

The 97 mm blower

97 mm blowers (San Ace B97, GDSTIME 9733 and clones) are the beginner default

@ju5t5omedude's beginner route (2026-08-04): a 4-pin Sanyo San Ace B97 blower (97 x 33 mm) rated at least 1.5 A, a knob or USB PWM fan controller, and either aluminium HVAC tape or a printed A100 bracket. It is what most members use, sold for example as a GDSTIME 12 V PWM 9733 two-pack or as "97mm blower 12v 4 pin" in 1.35 A and 1.85 A versions. Clones cost $5 to $12 on eBay and work, so no Delta or Sanyo brand is needed. The 97 mm blower is advertised as a puller, and a cardboard plenum is enough; other fans need a custom adapter, especially for a 180-degree turn. Results:

• A 1.8 A B97 capped at 60 % keeps a card at 68 C under a sustained 220 to 260 W; at about 4,200 rpm it holds around 71 C.
• One 1.85 A B97 on a USB fan controller kept a card below 65 C at 250 W sustained through hours of video generation. The curve had to be capped at 66 % to stay clear of the controller's overload protection, and he usually runs 40 to 50 %. A curve keyed to hotspot temperature hunted, so he may key it to die temperature.
• In an HP Z4 G4, two cards with HVAC-taped 1.85 A B97s, limited to 66 % by a USB PWM controller driven from CoolerControl, run LLM inference at about 65 C. Dual blowers hold 70 to 73 C core and can be kept under 70 C (2026-08-23).

Choosing and wiring one:

• The 1.85 A models are about 65 dB but lower pitched, closer to white noise than axial fans. 1.85 A has plenty of headroom for 300 W, 2.7 A is overkill, and he planned to drop to 1.5 or 1.2 A. Fan-controlled B97s are not bad at idle; the blower is loud but tolerable at low speed.
• The blowers plug into a USB controller rather than the motherboard. NZXT's v2 USB PWM controller (3 PWM outputs) with fan curves in CoolerControl works, and CoolerControl may also drive spare headers on third-party motherboards; @skyflyingfish also reports CoolerControl works well. For four stacked cards he suggests four 1.85 A B97s on a sturdy PWM controller.
• The Delta BFB1012VH comes as 1.2 A 3-pin, 1.8 A 4-pin with a GPU-style connector (that one would need a header soldered to the card) and 2.7 A 4-pin with a normal header. @redblood_'s BA10033B12H is 3-wire: it works but runs at 100 % all the time.
• Others noted that blowers may not fit most cases (see Fitting a blower in a case below).
• For other passive datacenter cards that ship with small 40 mm fans (a V100 in that thread), he listed BFB1012VH, PVB120G12N-P01 and 9BAM12P2G03 as quieter blowers. @endsjeremy: a 120 mm fan can only just hold such a card at full TDP and needs a large shroud.

Printed parts he linked: the Level1Techs A-series blower fan adapter thread, Thingiverse 7388824 ("cmp170 fan" by alexpon1997, an upgraded revision), a 180-degree duct for the B97 on Printables 1791789, a straight A100/A800 adapter on Thingiverse 6773548, and a teardown and centrifugal fan mod video (YouTube).

Proven
Source: @ju5t5omedude, 2026-07-24
Source: @farawaythrower, 2026-07-24
Source: @quark2world, 2026-07-24
Source: @replacedriver47, 2026-08-04
Source: @ju5t5omedude, 2026-08-04
Source: @ju5t5omedude, 2026-08-04
Source: @fallengod9962, 2026-08-12
Source: @ju5t5omedude, 2026-08-12
Source: @skyflyingfish, 2026-08-13
Source: @ju5t5omedude, 2026-08-15
Source: @redblood_, 2026-08-19
Source: @ju5t5omedude, 2026-08-19
Source: @_arcanum_, 2026-08-19
Source: @ju5t5omedude, 2026-08-23
Source: @stcord, 2026-08-23
Source: @con3po7016, 2026-08-23
Source: @ju5t5omedude, 2026-08-31
Source: @endsjeremy, 2026-08-31

A 97 x 33 mm blower: 72 C at 3,000 rpm, 80 C at 2,500 rpm

@vcore2191 measured full-load steady-state temperatures with a 97 x 33 mm blower fitted in a closed case: 72 C at 3,000 rpm and 80 C at 2,500 rpm. Adding a 120 mm 1,200 rpm axial fan with a sealed funnel at the card outlet lowered the temperature to 78 C with almost no added noise, while two extra 120 mm case fans aimed at the card gained only about 0.5 C more. Looking at a printed 2x 120 mm to 4-card intake plenum posted by @stcord, he noted that axial fans only make enough static pressure at very high rpm (4,000 in that design), which is still audible from the next room.

Reported
Source: @vcore2191, 2026-08-23
Source: @stcord, 2026-08-23

Blowers hold 60 to 67 C at 175 to 200 W

• @froggao: a Delta BFB1012EH PWM blower on a printed bracket with high intake airflow, 67 C max, not loud.
• @68x0: the PluraliT integrated housing with a PLB08025B12HH blower is loud at full speed, but at 50 to 60 % duty keeps cards at 60 C under heavy load with a 200 W power limit. @trustinter: the PluraliT print lists two blowers, PLB08025B12HH (2.5 A) and CF7525U12D (1.2 A).
• @xtj7: an 18 W 9733 at 20 % duty holds 60 C at 175 W and is quiet at that speed.
• @_orrorin runs 9733 blowers on a fan controller with curves built with an AI assistant; the cards stay cool and the noise stays low in a small apartment.
• @kellertreppe123 runs a 3 A blower per card at a steady 60 C under load and calls it fairly quiet (@zorg33: almost as loud as a server axial fan). He stresses that a datacenter GPU needs air blown into it. For 180-degree duct adapters he says splitting the funnel in half with a small fin improves the airstream and claims 30 to 50 % more air pressure; that figure was not measured.
• @prefix10k reported insufficient cooling with blower fans (design not given).

Proven 8 GB and 10 GB
Source: @kellertreppe123, 2026-09-25
Source: @prefix10k, 2026-10-06
Source: @wakeupmbrc1, 2026-10-15
Source: @froggao, 2026-10-15
Source: @68x0, 2026-10-15
Source: @xtj7, 2026-10-15
Source: @trustinter, 2026-10-15
Source: @kellertreppe123, 2026-10-20
Source: @_orrorin, 2026-10-20
Source: @zorg33, 2026-10-20

How big a blower

Minimum blower sizing: 0.47 A is the floor, a 1 A 9733 cools one card to 250 W

• @_arcanum_: a Nidec G10D 0.47 A blower is the bare minimum for 250 W sustained, and also the most noise he tolerates next to his desk. He thinks the AliExpress direct-mount blower @fff7d1bc planned to buy is the same Sunon he has: borderline at 100 % in a 25 C air-conditioned room, maybe usable on repasted cards.
• @matstatman runs the U-shaped 0.47 A version and sees 76 C, which he calls too hot and very loud. @ju5t5omedude noted that bracket looks made for a Dell dual-inlet blower slightly narrower than the B97.
• One owner's 0.9 A Sunon blower (about 28 CFM) held a card at full load to about 65 C after an hour.
• @ob88594: the fan must draw more than 8 W to hold 250 W, and any 9733 rated above 7 W keeps the card under 80 C (2026-08-22). On 2026-09-15: most 97 x 97 x 33 blowers rated 12 V 1 A cool one card acceptably up to 250 W.
• @xtj7 bought a 1.5 A 9733 for about $8 in case a Chinese "1 A" rating is optimistic. On a lab power supply it drew 1.05 A at full speed (wired with only + and -, so no PWM), and he thinks it is overkill at a planned 175 W with a custom shroud that seals well. @gpmickey also uses 1.5 A fans: the Amazon listing said 1.8 A, the housing says 1.5 A.

Proven
Source: @91d5a66bffc978331fac75b68fa86aba, 2026-08-15
Source: @ob88594, 2026-08-22
Source: @_arcanum_, 2026-08-22
Source: @fff7d1bc, 2026-08-22
Source: @matstatman, 2026-08-22
Source: @ju5t5omedude, 2026-08-22
Source: @ob88594, 2026-08-22
Source: @ob88594, 2026-09-15
Source: @xtj7, 2026-09-15
Source: @gpmickey, 2026-09-15

Small 3 W blowers: 250 W for one owner, not for another

Contested

• @_arcanum_: his small printed blower adapter with an about 3.3 W blower holds 250 W sustained at 80 % PWM at about 70 C core in 25 C ambient, on a card that was not repasted, quiet enough to sit next to with earphones. Source: @_arcanum_, 2026-08-05
  Source: @rubbra_f, 2026-08-05
  Source: @ju5t5omedude, 2026-08-05
• @quark2world, with what he believes are the same 3 W blowers and adapter, cannot keep a sustained 250 W below 80 C at about 25 C ambient. Whether mounting, card or paste made the difference was not resolved. Source: @quark2world, 2026-08-05
• Later @quark2world runs a card permanently at a 180 W power limit with a 3 W blower at about 2,500 rpm in a printed shroud sealed with electrical tape: around 67 C after hours of work, touching 70 to 72 C on spikes. An 80 mm Noctua Redux puller fan lets the blower idle at low rpm (at 1,000 rpm the blower alone lacks pressure), and a third fan clears hot air that the SP3 CPU cooler blows at the card. A custom script sets the fans over IPMI from the card temperatures; at idle it sounds like a laptop. This replaces his earlier view that a 3 W laptop blower is only fine for functional testing. @_studebaker_ independently found about 175 W to be the point where the card stops constantly hitting the power limiter; @ju5t5omedude uses Dell blowers rated 1.2 A and 1.5 A. Source: @quark2world, 2026-08-22
  Source: @_studebaker_, 2026-08-22
  Source: @ju5t5omedude, 2026-08-22

BC250-style 12 W PWM blowers (about $6): untested

Contested

• @phoenixone7 ordered a pair of 12 W PWM blowers, about $6 each, that people used successfully on BC250 boards; @ju5t5omedude thought they should be plenty. Source: @phoenixone7, 2026-08-04
  Source: @ju5t5omedude, 2026-08-04
• Pasteyy: BC250s do not reach 300 W and their coolers are less dense, so less restrictive. Phoenix17 countered that a BC250 can pull about 220 W through fairly dense fins. Source: @pasteyy., 2026-08-04

No result was posted; he later used 1 A blowers (see Phoenix17's 1 A blower adapter below).

Warning: fake current ratings on cheap blowers (real draw about 0.4 to 1.2 A)

Warning Labels on cheap AliExpress blowers overstate the current, so a "high-current" blower may be weak. @skirtchaser bought several BFB0712HF fans rated 1.8 A for MI50s and all were fakes to varying degrees: real draw 0.4 to 0.7 A, peaking around 1.2 A with the airflow almost fully blocked. Even "WINSINN" 120 mm blowers from the WINSINN Official Store turned out fake, and most AliExpress fans he received were fakes of some kind. @squishysquare: cheap AliExpress fans overstate their current, real full speed being about 1 to 1.2 A. @ju5t5omedude and @dalzyyy doubted the 3 A label on DaLZy's own blower, and MI50-style fans sold under one 75 mm part number range from 72 to 76 mm. What to do instead: size by measured draw, not the label (a "1.5 A" 9733 drew 1.05 A on a bench supply; see Minimum blower sizing above), and check the frame size before you print an adapter for it.

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

Seller-bundled fans with a GPU-style connector were too weak (85 C in memtest)

@matstatman's card came with a fan on a small GPU-style connector that does not fit motherboard headers. @ju5t5omedude: use a converter pigtail (a mainboard fan extension cable, checking that the wire colours match) or solder on the pigtail from a dead 4-pin case fan, and expect the bundled fan to be low amperage. Wired straight to 12 V (full speed), the card unlocked, but it reached 85 C in memtest_vulkan with fan-limited throttling: too weak for the card.

Reported
Source: @matstatman, 2026-08-14
Source: @ju5t5omedude, 2026-08-14

Builds with numbers

Four cards in a tower with $2 blowers: 50 dB at 60 %, 76 C at 250 W

@ob88594 built a 4-card system in a Cooler Master HAF II 500 case, made to look like a normal gaming PC. The blowers are $2 turbine fans between 9733 and 12032 size, cut to fit with a saw and taped on with high-temperature (Kapton) tape. They run from an external 12 V fan speed controller, and he had a Pico RP2040 on order to set fan speed from nvidia-smi temperature and load readings.

• All four fans at full speed, 250 W per card: peak 76 C, 55 dB at 50 cm measured with a phone (under 56 dB).
• 60 % fan speed, 200 W TDP: 75 C at a normal 50 dB.

He says 9733-and-larger blowers make a non-violent noise close to a normal PC, unlike small high-rpm fans, and larger fans mean lower power and noise.

Reported
Source: @ob88594, 2026-08-22
Source: @ob88594, 2026-08-22

A Dell OptiPlex 7060 blower (BUC1412VE-00) taped on, no printer needed (contested at 250 W)

Contested

• @ob88594: the Dell OptiPlex 7060 fan (BUC1412VE-00) is quiet and effective and needs no adapter bracket. The fins need very high air pressure, so saw off the upper part of the fan's top cover and tape the blower to the tail of the card. It is only worth it at $5 or less (he paid $2 each); otherwise use a 9733 blower rated above 7 W, with adhesive or a printed interface. 40 % fan speed handles long inference loads quietly, and 30 minutes of gpu-burn held 75 C at 70 % PWM, with the HAF II 500's case fan adding airflow. Source: @ob88594, 2026-08-22
  Source: @ob88594, 2026-09-15
  Source: @gpmickey, 2026-09-15
• @sojun80 disputed the result: the cards in the screenshot drew only about 150 W, so he asked for a test with video renders or mining. Source: @sojun80, 2026-09-15

Four stacked cards: Nidec 9733 blowers at 1,800 rpm keep 180 W cards under 80 C

@syntheticblocks runs four stacked cards with 9733 blowers (Nidec G10D12BS8AZ-57Z99) on printed shrouds. One fan is stuck at 2,700 rpm (audible); the other three at 1,800 rpm keep the cards under 80 C at a 180 W power limit with tolerable noise. The top card runs coolest and the bottom card hottest. The fans are controlled in software with CoolerControl. @68x0 recommended the Arctic USB fan hub (Linux compatible), which Syn already uses for case fans, but the Nidec connectors are too shallow to plug in easily; 68x suggested a pigtail.

Reported
Source: @syntheticblocks, 2026-09-16
Source: @68x0, 2026-09-16

A Tesla P100 blower taped on: 220 W with HBM at 75 C

@wakeupmbrc1 cools a card with a Tesla P100 turbine (blower) taped to it. At 220 W peak the HBM reaches 75 C, with noise "not too loud but not silent" (2026-08-30). On 2026-09-02 he said the P100 turbine at 65 to 70 % speed is enough for a 220 W card.

Reported
Source: @wakeupmbrc1, 2026-08-30
Source: @wakeupmbrc1, 2026-09-02

120 mm blowers: under 60 C benching, 72 C core and 75 C HBM at 300 W

@letterc67 asked whether a 120 mm blower beats a 97 mm one (no adapter existed yet) and proposed the Delta BFB1212VH: 12 V, 1.25 A rated, up to 1.88 A depending on the variant, about 15 W (datasheet).

• @jimtjames already used a 120 mm blower on a makeshift adapter: about 72 C core and 75 C HBM under full load at 300 W, with a water block planned for real use. @ju5t5omedude noted that shorter adapters can use the blower's mounting holes as an L-bracket with pegs, and @meromgreen designs printed parts by having GPT write OpenSCAD files from caliper measurements.
• roboalchemist runs 2-pin 12 V 120 x 32 mm dual-ball-bearing blowers (the listing claims about 3,000 rpm) at constant full speed and never exceeds 60 C while benchmarking. Behind a door he cannot tell whether it is on, unlike the stock 40 mm fans that were audible through two doors. @vcore2191 noted the bench power was low and a Vulkan memtest would load the cards harder. roboalchemist was also testing 97 x 97 x 33 Sonew blowers to fit inside a closed case.
• @ju5t5omedude: a 4-pin PWM 120 mm blower is the PMB1212PLB2 (search for "120mm x 32mm 4 pin blower"), and spec sheets suggest it is about 10 dB quieter than 97 mm blowers. An STL for a 120 mm radial fan is Thingiverse 5293965.

Pointer: roboalchemist's adapter (CMP170HX_Blower_-_CMP100_Side_Fan_-_blower_medium.stl) is in the Discord; see Discord. It is not re-hosted here until the author agrees.

Reported
Source: @letterc67, 2026-08-12
Source: @jimtjames, 2026-08-12
Source: @ju5t5omedude, 2026-08-12
Source: @meromgreen, 2026-08-12
Source: @roboalchemist., 2026-08-23
Source: @vcore2191, 2026-08-23
Source: @ju5t5omedude, 2026-08-23

Four unlocked 8 GB cards in a mid-tower with taped 120 mm blowers: 65 C core, 75 C HBM

@unmotivatedgene built four unlocked 8 GB cards into an old mid-tower gaming PC with one 120 mm blower per card, held on with gaffer tape. The fans run from motherboard fan headers, each tied to its card's temperature (a manual controller had forced high speed all the time). A 15-minute DeepSeek V4 Flash burn-in on three cards peaked at 65 C core and 75 C HBM, and idle noise is like a normal gaming PC. The fans are lined up on a tower of skewers or dowels with velcro spacers and splayed slightly apart, away from the cards, because tight spacing restricts airflow. The side panel closes flush, and the fourth card sits on a riser.

Reported 8 GB
Source: @unmotivatedgene, 2026-08-23

Other blower builds: a Dell blower, and a custom shroud at 82 C and 250 W

@ju5t5omedude tested a Dell blower as a possibly quieter alternative to the B97 that would not overload fan headers; the shroud needs trimming and plugging. @petruswang built a custom blower shroud that reached 82 C at 250 W in 30 C ambient ("not so good"). @dalzyyy advised more tongue clearance, a larger outlet and moving the tongue forward, and @quark2world suggested a power limit.

Reported
Source: @ju5t5omedude, 2026-08-19
Source: @petruswang, 2026-08-19
Source: @dalzyyy, 2026-08-19
Source: @quark2world, 2026-08-19

Printed blower coolers and adapters

Public models are linked; files posted only in the Discord get a pointer and wait for their authors. Every printable file is indexed on Printable shrouds and ducts.

DaLZy's 80 mm blower cooler: design and where to get it

@dalzyyy designed a full replacement shell that keeps the stock copper heatsink, which covers nearly the whole PCB, and adds an 80 mm blower. The stock aluminium shroud cannot be modified without CNC, and an internal 6 to 7 cm blower inside the stock shell would cool about 150 W at most. He used the Fusion MCP server with Opus to design the tongue and volute.

• Where to get it: published on 2026-08-05 as "80mm Blower Cooler for 170HX/A100" on MakerWorld (model 3138171) and Printables 1801217, with thermal results on those pages; @_j3ns pinned it in the Discord.
• Fan: it fits only the Champion CF8038U12D blower (other fans need model edits), sold on Taobao as a "4090 turbine fan". It is hard to source outside China; search for "4090 blower" (@_arcanum_). DaLZy said a fan of about 72 mm would suit the design better, and in hindsight a 66 mm blower would have been easier because the EPS cable gets in the way of an 80 mm one.
• The shroud replaces the original rather than sitting over it. The backplate mainly serves as the fan mount and lets air flow over the back of the PCB. @ju5t5omedude worried about airflow to the PCB rear; DaLZy said it may matter when stacking cards.
• Printing: the first model needed a 270 x 270 mm bed plus supports. The version with modelled supports needs a 220 x 230 x 235 mm print volume, and on 2026-08-10 he said those modelled supports did not work yet and should be removed before slicing.
• Results: it performs very well at 250 W and can handle 300 W+ if you accept more noise (2026-08-27); about 75 C at 60 % fan speed (2026-09-16). @port00 was modifying it for better printability and fan compatibility. DaLZy had also shown a new blower on 2026-08-22 that he hoped would handle 200 to 250 W quietly.
• Fan alternatives raised: the BASA0725R2U blower used by the MI50 (gfx906) community (@aaron0010; DaLZy doubted it has enough power), and Gigabyte blower-card 80 mm fans at about $15 (@sealclubber).
• Before printing any blower shroud, check that you can buy the fan it was designed for (@quark2world). @schakaqt: temperatures depend on the fan; an $8 5,000 rpm AliExpress blower is very annoying, while an Arctic 4K server fan at 3,000 to 4,000 rpm is bearable. @darcidride compared it with an Nvidia A2 customizable server-GPU shroud (MakerWorld 1372280) and asked for a pinned summary of 2x, 3x and 4x shrouds.

Pointer: DaLZy's source files (Shroud.step and Backplate.step) were posted in the Discord; see Discord. Use the public MakerWorld and Printables pages; the Discord files wait for the author.

Reported
Source: @dalzyyy, 2026-08-01
Source: @tugoubxs, 2026-08-01
Source: @ju5t5omedude, 2026-08-01
Source: @gmzombie, 2026-08-01
Source: @dalzyyy, 2026-08-05
Source: @dalzyyy, 2026-08-05
Source: @cute_sc2, 2026-08-05
Source: @aaron0010, 2026-08-05
Source: @sealclubber, 2026-08-05
Source: @robbbbbb7973, 2026-08-06
Source: @_j3ns, 2026-08-06
Source: @rubbra_f, 2026-08-06
Source: @letterc67, 2026-08-06
Source: @dalzyyy, 2026-08-09
Source: @ju5t5omedude, 2026-08-09
Source: @mountains1605, 2026-08-11
Source: @_arcanum_, 2026-08-11
Source: @darcidride, 2026-08-20
Source: @schakaqt, 2026-08-20
Source: @quark2world, 2026-08-20
Source: @replacedriver47, 2026-08-20
Source: @dalzyyy, 2026-08-27
Source: @port00, 2026-08-27
Source: @mfj4, 2026-08-27
Source: @dalzyyy, 2026-09-02

DaLZy's blower cooler: measured (gpu-burn 297 W at 71/73 C core, 72/74 C HBM)

Revision 0.1 (2026-08-04) was a two-piece clamshell around an 80 mm blower sold as a "4090 blower" (maximum about 5,300 rpm), 302 mm long against about 267 mm for the stock card. At 2,500 rpm it could not sustain 300 W and hit an 85 C wall. Two 10-minute gpu-burn runs:

Run	Fan	Average power	Core (avg/max)	HBM (avg/max)	Noise
1	maximum	297 W	71/73 C	72/74 C	over 75 dBA; the cooler vibrated at its resonant frequency
2	2,950 rpm	249 W	76/81 C	77/82 C	about 55 dBA

He suspected his thermal paste application, and thought a 66 mm fan could bring the length down to about 280 mm. At 2,950 rpm he called it "kinda loud but acceptable", and he expects it to handle a 150 W limit at 2,500 rpm without being loud, depending on the room (2026-08-05). Later noise figures: about 60 dB at 30 cm at 70 % PWM, measured with an Apple Watch (2026-08-12), and an estimated 50 dB at 30 cm at 3,000 rpm (2026-09-13). @skyflyingfish found blower noise too loud and was considering water cooling; @lenxtroz planned to ask DaLZy for a quote on this fan mod.

Reported
Source: @dalzyyy, 2026-08-04
Source: @dalzyyy, 2026-08-05
Source: @skyflyingfish, 2026-08-12
Source: @dalzyyy, 2026-08-12
Source: @dalzyyy, 2026-09-13
Source: @spiralglide, 2026-09-13
Source: @lenxtroz, 2026-09-15

DaLZy's blower cooler: build problems, printing tips and member mods

• Cracked fan: DaLZy's blower cracked after 2 days of load (2026-08-08). He thinks it was squeezed in shipping, with the shroud's low tongue clearance adding pressure on the fan; @ju5t5omedude suspects rubbing or buffeting from a surface that is not flat. The fan is cheap, so @letterc67 planned to order spares.
• Assembly: aligning the fan to the screw holes is hard, so @meromgreen suggested a cutout for the fan base, which DaLZy added. Thread the wire through the slot first, fit the screw opposite the slot, then fit the other two.
• Other fans: Merom Green test-fitted a BASA0725R2U blower (about 75 mm, same mount spacing). The diameter fits, but it is shorter, leaving a large gap at the lip and needing longer screws, so he suggests the Taobao "4090 turbo" fan. DaLZy considered a redesign around an RTX 4070 FE fan (AliExpress item 1005008281102516). @ju5t5omedude prefers the BASA because it allows more taper and card spacing, and suggests copying V620 and MI50 shroud designs.
• Printing the backplate (2026-08-09): Merom Green's 45-degree print failed at 80 to 90 % on a Bambu X2D (Bambu PETG translucent and PETG basic, bed 70 C, chamber 30 to 40 C), because the thin section is only 5 layers at 0.2 mm and flexes with XY movement. DaLZy prints it at 45 degrees on a Qidi Q2 with a little warping and prefers that angle for finish and strength, but cannot thicken it much without losing the two-slot profile. Merom Green's fix: split the backplate in half and print it flat with a PLA support interface, about 2 hours against about 9 hours with full supports; the backplate needs little strength because the front shroud carries the load. @mountains1605 planned to print it in PETG-CF for four cards.
• @aaasss1456 built a modified version: he set the gap between the fan and the top cover with the fan screws to stop rubbing, sealed it with Teflon tape and added temperature-controlled fan speed.

Reported
Source: @dalzyyy, 2026-08-08
Source: @meromgreen, 2026-08-08
Source: @meromgreen, 2026-08-08
Source: @letterc67, 2026-08-08
Source: @ju5t5omedude, 2026-08-08
Source: @dalzyyy, 2026-08-08
Source: @ju5t5omedude, 2026-08-08
Source: @meromgreen, 2026-08-09
Source: @dalzyyy, 2026-08-09
Source: @mountains1605, 2026-08-10
Source: @aaasss1456, 2026-08-10

Phoenix17's 1 A blower adapter, with PWM through CoolerControl

@phoenixone7 fitted 1 A blowers, about $6 each on AliExpress and chosen for cost, current draw and noise, with a printed adapter. He wrote the part number as BUC10212VJ-00 in one message and BUC1012VJ-00 in the next, so check his photos. Blower section: external 87.6 x 25 mm, internal 83.8 x 20.5 mm, outlet about 87.5 x 25 mm, close to the card's 92.5 x 23.5 mm inlet.

• A first crush-fit adapter (no taper, minimal length) was blocked by the power connector latch. An extended version fit, and a triple-inlet version that also covers the side inlets fits too but needs print supports.
• The posted adapter is very snug and may not need tape. It makes the card about 47 cm long (@quark2world); Phoenix17 said "the cooler itself" can probably be cut. @ju5t5omedude noted the 97 mm blower is the tried and true alternative.
• Results (2026-08-26): with 4-pin PWM extension cables, so that CoolerControl drives one blower per card, both cards stayed under 80 C hotspot in testing at 250 W. The shroud stays attached and cools the memory well, and the 1 A blowers lack the usual server-fan whine even at high speed.

Pointer: the adapter (170hx_duct_over_v13.stl) is pinned in the Discord's 3D-prints channel; see Discord. It is not re-hosted here until the author agrees.

Reported 8 GB
Source: @phoenixone7, 2026-08-11
Source: @_j3ns, 2026-08-11
Source: @meatsus, 2026-08-12
Source: @ju5t5omedude, 2026-08-12
Source: @phoenixone7, 2026-08-12
Source: @_j3ns, 2026-08-12
Source: @quark2world, 2026-08-12
Source: @phoenixone7, 2026-08-26

A Sanyo 9BMB adapter, an LLM-made funnel, and a stackable 9733 adapter

• @civiloid made a printable mount for the Sanyo 9BMB12P2G01 (97 mm) blower. At 35 % speed it already cooled the card more than enough, but below about 45 % it vibrated badly, and these fans are not quiet at low rpm. He says the adapter worked well enough, but the blower is expensive and too loud even for a basement machine, so he moved to water cooling (see Noise below). He produced the mount by giving an LLM ("GPT 5.6 sol") the reduced card-side STL and the maker's fan STL and iterating. His 2026-09-02 post says it took 18 iterations and a second print (screw hole size mismatch), with the cable cutout the hardest part; a 2026-10-13 post says about 5 attempts. @reala7096 also used that model to modify an STL: not very good, but it gets there with persistence.
• @kellertreppe123 also generated a funnel with an LLM: the latches ended up on the wrong side, and the funnel is split diagonally for throughput.
• Stackable 9733 adapter: @syntheticblocks shared an online Level1Techs A100 blower adapter that Astra modified with a larger mouth and lower overall height for tightly stacked cards (2026-09-16). The mouth may be a fraction of a millimetre small for some 9733 fans. @cressida1881 printed it and it fit after a slight adjustment with a hot-air station; it is slightly long but fits a normal case. @fafinet was still looking for a stackable 9733 adapter in October, since most files online are for spaced cards or axial fans on 4 cards.

Pointer: civiloid's adapter (cmp170hx_sanyo9bmb_adapter_v18.stl and the source zip) is in the Discord; see Discord. It waits for the author.

Pointer: Syn's stackable adapter (l1_a100_blower_stackable_power_latch_original_fan.stl), derived from the public Level1Techs A100 blower adapter, is in the Discord; see Discord. It waits for the author.

Reported 8 GB and 10 GB
Source: @civiloid, 2026-08-25
Source: @justarionlyari, 2026-08-25
Source: @civiloid, 2026-09-02
Source: @reala7096, 2026-09-02
Source: @ju5t5omedude, 2026-09-02
Source: @syntheticblocks, 2026-09-16
Source: @cressida1881, 2026-09-16
Source: @civiloid, 2026-10-13
Source: @fafinet, 2026-10-13
Source: @kellertreppe123, 2026-10-13

Fitting a blower in a case

Length (over 40 cm), clearance and tools

• A card with a blower attachment is over 40 cm long and blocks mATX front-panel headers, so @quark2world calls this not a beginner build. He cannot recommend his own blower adapter type: there is no room for a screwdriver and the EPS/PCIe power adapter gets in the way, though it does cool the card.
• @ju5t5omedude fits two blower-cooled cards in an HP Z4 G4, using a tape adapter for clearance, and estimates four would fit in a Qube 540 case. He finds the 20 W+ B97s loud and overkill; a USB fan controller helps, and a 12 W Dell CPU-cooler blower part works.
• Shorter options: @artistosu uses a 2x 40 mm fan shroud (68 CNY with fans, 45 mm long, good for SFF cases; Cults3D) with a rear 120 mm Arctic P12 Max in push-pull. @moliminous suggested a shroud as used with Tesla P40s, plus a server chassis.
• Printing your own: @meatsus says a second-hand Bambu A1 (about 150 EUR) prints countless ducts, and paid print services cost too much. @lubakmanqk has two Bambu printers and planned to test ducts.

Reported
Source: @quark2world, 2026-07-31
Source: @meatsus, 2026-07-31
Source: @lubakmanqk, 2026-07-31
Source: @ju5t5omedude, 2026-08-03
Source: @quark2world, 2026-08-03
Source: @artistosu, 2026-08-03
Source: @moliminous, 2026-08-03
Source: @redblood_, 2026-08-03

Noise

Blowers have two modes, "off" and "violence"

• @civiloid's high-rpm Sanyo 9BMB12P2G01 blowers at 70 % made a low-pitched noise that escaped a basement compartment through three closed doors and concrete walls (audible in the elevator). At 60 % they resonated and carried worse even as the dB figure dropped; at 35 to 40 % the sound was acceptable, but they could no longer sustain 200 W per card. Two 40 mm fans per card were not enough for him either, partly because the basement air got warm, and a 2x 120 mm Arctic 4,000 rpm setup was quiet enough but could not cool all four cards. He moved to Water cooling.
• @matej87: blowers without rpm control are very loud, and water blocks add about $250 per card for quiet operation. @voidgiraffe was dreading the noise before installing his cards; he now runs eight blowers for four cards, and above 50 % it is painful.
• @fff7d1bc: the options seem to be blowers with two modes, "off" and "violence", or big fan shrouds, with nothing good for a quiet two-card setup.
• @skirtchaser: 8-GPU servers have a bigger noise problem than mining-rig style builds, even with blowers.
• @polskiekaktus linked a test of a 200 mm two-stage blower reaching 8.7 mm of static pressure at 47.4 dBA (YouTube).

Other noise figures on this page: the 1.85 A B97 at about 65 dB, a 4-card tower at about 50 dB at 60 %, and DaLZy's cooler at about 55 dBA at 2,950 rpm.

Reported 8 GB and 10 GB
Source: @matej87, 2026-08-25
Source: @voidgiraffe, 2026-08-25
Source: @voidgiraffe, 2026-09-01
Source: @fff7d1bc, 2026-09-01
Source: @civiloid, 2026-09-01
Source: @skirtchaser, 2026-09-01
Source: @polskiekaktus, 2026-09-01
Source: @civiloid, 2026-10-13

Related

Fans and ducts, Printable shrouds and ducts, Cooling and power, Thermal limits, Undervolting and power limits, Water cooling, Chassis builds, Monitoring, Testing and burn-in, Thermal interface: paste, pads and phase-change.

Imported from the cmpunlocker Discord with credit; see Credits.