Fans and ducts
From CMP Wiki
More actions
The CMP 170HX has no fan of its own. You add one, and you make sure its air goes through the heatsink fins and not around the card. This page collects what people run, with the numbers they posted. Start with Cooling and power if you have not read it.
What works, in short
- Put a fan on the intake end of the card (the end opposite the slot bracket) and seal the path so the air has to go through the fins. Open gaps waste most of the airflow.
- Use fans with real static pressure. A 120 x 38 mm server fan (Arctic S12038-4K) is the community favourite.
- Push-pull (a second fan pulling at the bracket end) lets the fans run slower for the same temperature.
- Drive the fans from the hottest card's core or HBM temperature, not from the motherboard CPU curve.
- The card's own fan header is useless (no PWM, no fan curve in the VBIOS). Use the motherboard, a hub or a controller.
Measured setups
Temperatures are the hottest sensor the poster gave. "Cap" is the power limit set with nvidia-smi -pl.
| Setup | Cards | Cap | Result | Noise | Status | Source |
|---|---|---|---|---|---|---|
| One Supermicro FAN-0074L4 80 mm fan per pair of cards, taped (VHB) straight onto the card ends, no duct | pairs | 185 W | never above 63 C, inference with 1M-token contexts | "not soul-piercing" | Reported | Source: @c133, 2026-10-09 |
| One Arctic P12 Pro 120 mm per pair, four cards on a 5-slot PCIe switch with the middle slot left empty | 4 | not given | "pretty happy" with the airflow | low | Reported | Source: @cibernox, 2026-10-07 |
| One 140 mm fan for three cards and one 120 mm fan for two cards, printed shroud | 5 | not given | "no temp issues" | not given | Reported | Source: @analyticeth, 2026-10-07 |
| Arctic P12 Pro 120 mm on a printed shroud | 1 | 175 W | 65 C under ComfyUI | not given | Reported | Source: @rrickk420, 2026-10-07 |
| Air cooling on a 64 GB (unlocked 8 GB) card | 1 | 175 W | 60 C core, 65 C HBM | not given | Reported | Source: @rrickk420, 2026-10-08 |
| Sealed high-angle printed funnel, fan sealed with VHB and gaffer tape | 1 | 150 W | 70 C; lots of blowback in the funnel | not given | Reported | Source: @globalapathy, 2026-10-07 |
| Arctic S12038-4K, two push and two pull | not given | 300 W | HBM 72 C at 3,500 rpm in a 24 C room | "pretty happy" | Reported | Source: @wakeupmbrc1, 2026-10-16 |
| Two S12038-4K per card in push-pull, plus three as case intake, no grills | not given | not given | no temperatures posted | "hear basically nothing" | Reported | Source: @froggao, 2026-10-26 |
| S12038-4K push only, fans regulated on the hottest core or HBM of all four cards | 4 | 200 W | at most 62 C, in a cold garage | not given | Reported | Source: @alanoo, 2026-10-26 |
| Two S12038-4K at 3,500 rpm for four cards, pull fan standing on the floor at the exhaust | 4 | 250 W | "more or less enough", hotter than wanted | not given | Reported | Source: @civiloid, 2026-10-16 |
| Two-fan plenum (remix of the Printables plenum) with two S12038-4K, about half speed all day | 4 | not given | fans go above 2,000 rpm only under sustained agentic load | not given | Reported | Source: @mountains1605, 2026-10-08 |
| Three-card plenums with Arctic P12 Pro in push and pull, airflow balanced by partly blocking the cooler cards | 6 | full power, 10 minutes | hottest card settles at 80 C | not given | Reported | Source: @reala7096, 2026-10-08 |
| Three push and three pull Delta server fans at 100 %, no plenum yet | 8 | 175 W | hottest card 70 C, the rest 60 C | "white noise" level | Reported | Source: @neapopa, 2026-10-22 |
| Two Arctic P14 Pro in push-pull at 40 to 50 % (about 1,500 rpm), cardboard exhaust shroud, gaps foamed, in an ATX case | 3 | 120 W (spikes 130 to 160 W) | 69 to 79 C under test load; "not for the summer" | airflow and VRM chirp | Reported | Source: @pekoraali, 2026-10-19 |
| Eight 80 mm fans in push-pull | 8 | not given | not given | "very quiet" | Reported | Source: @con3po7016, 2026-10-08 |
| Printed integrated housing (PluraliT) with a PLB08025B12HH blower at 50 to 60 % | not given | 200 W | 60 C under heavy load | loud at full speed | Reported | Source: @68x0, 2026-10-15 |
| One 18 W 9733 blower at 20 % | 1 | 175 W | 60 C "easily" | screams at full speed | Reported | Source: @xtj7, 2026-10-15 |
| Delta BFB1012EH PWM blower on a printed bracket | 1 | not given | 67 C max | "not loud" | Reported | Source: @froggao, 2026-10-15 |
| One 3 A blower per card | not given | not given | steady 60 C under load | as loud as a server axial fan (Source: @zorg33, 2026-10-20) | Reported | Source: @kellertreppe123, 2026-10-20 |
| AMD V620 80 mm shrouds (almost fit; a heat gun would fix it) with Arctic S8038-10K fans, HVAC tape | not given | not given | "temperature is not a problem"; 50 % speed is fine for LLM loads | heard across the house at full speed | Reported | Source: @unlikelypotato, 2026-10-06 |
| Two 40 mm AVC DBTB0428B2G axial fans per card | not given | not given | under 75 C | piercing drone at 30 to 40 % PWM on light loads | Reported | Source: @fafinet, 2026-10-13 |
| Two 40 x 40 mm fans from an Antminer PSU (about 16,000 rpm), card kept to two slots | 1 | not given | never above 60 C | unusable in the same room | Reported | Source: @shudza, 2026-10-08 |
| San Ace (Sanyo) 9BMB12P2G01 blowers on a printed adapter | 4 | 200 W | cooled the cards; at 35 to 40 % they could no longer hold 200 W | at 70 % audible through concrete walls; resonated below 70 % | Reported | Source: @civiloid, 2026-10-13 |
| Arctic S12038-8K at 20 to 25 % with a printed finger grill | not given | not given | not given | not given | Reported | Source: @givisok, 2026-10-26 |
| Printed integrated fan housing by RazK (slim version), on a 5-GPU PCIe switch board, high case airflow | 5 | not given | average 33 to 37 C, close to the owner's RTX 3090s | quiet unless fans are at 100 % | Reported | Source: @moliminous, 2026-10-10 |
| Supermicro fan at 50 % pointed roughly at the card, no shroud | 1 | not given | "good airflow" | not given | Reported | Source: @programmerboi, 2026-10-25 |
| 2U server, 3 A 80 mm fans | 1 | 250 W | not enough until the gridded slot brackets were taped over, which took 10 C off and made it stable | 4 such fans can draw 200 W | Reported | Source: @mlx4, 2026-10-20 |
| Printed shroud with the supplied Delta AFB0912H 92 mm fan (0.3 A) | 1 | 150 W | 80 C and above (air gap at the card and the cable; fan likely too weak) | not given | Reported | Source: @sh.ad, 2026-10-07 |
| Noctua industrial fans on fan controllers, cards on a mining rack with riser cables | 7 | 150 to 180 W | not given | under load like a hair dryer running nonstop; fine at idle; louder than RTX PRO 6000 Max-Q blowers | Reported | Source: @munisix, 2026-10-25 Source: @munisix, 2026-10-25 |
What the table shows:
- Proven One decent fan per pair of cards is enough at 150 to 185 W when it is pressed against the card ends. Proposed by @cibernox Source: @cibernox, 2026-10-07; reproduced with 80 mm Supermicro fans by @c133 Source: @c133, 2026-10-09 and with 120/140 mm fans by @analyticeth Source: @analyticeth, 2026-10-07. @programmerboi disputes that it is enough for sustained dense training, where he runs 9x Supermicro FAN-0163L4 (about 8 times the static pressure); both agree inference and training are different duty cycles. Contested Source: @programmerboi, 2026-10-09
- Proven Arctic S12038-4K fans in push-pull cool these cards well at moderate noise: 300 W per card at 3,500 rpm with HBM at 72 C Source: @wakeupmbrc1, 2026-10-16; two per card in push-pull "hear basically nothing" Source: @froggao, 2026-10-26; push only, four cards at 200 W, at most 62 C Source: @alanoo, 2026-10-26.
- Proven On the printed plenums an S12038-4K class fan is enough for inference, and Arctic P12 Pro fans alone do not hold a three-card shroud; use the 8K version for a full thermal stress test. Source: @reala7096, 2026-10-08 Two S12038-4K run four cards at about half speed all day. Source: @mountains1605, 2026-10-08
- At 250 W and above you need push-pull or a lot of fan. @zorg33 asked for 250 W solutions after modelling 2 to 3 times the airflow of 200 W. Theory Source: @zorg33, 2026-10-20
Fan models
Static pressure and current are the manufacturer's numbers where a link is given; the rest is what posters said.
| Fan | Size | Speed | Static pressure | Current | Notes |
|---|---|---|---|---|---|
| Arctic S12038-4K | 120 x 38 mm | up to 4,000 rpm | 11.45 mmH2O | 0.33 A (0.60 A max) | spec; 106 CFM. "Best noise/perf", "mogging any Noctuas, T30" Source: @wakeupmbrc1, 2026-10-26; about 400 g each Source: @wakeupmbrc1, 2026-10-26; the minimum the Printables plenum author recommends. Runs on a decent controller at 0.6 A Source: @port00, 2026-10-19 |
| Arctic S12038-8K | 120 x 38 mm | 800 to 8,000 rpm | 37 mmH2O | 2.2 A (3.2 A max) | spec; 220 CFM. For full thermal stress tests Source: @reala7096, 2026-10-08; needs a finger guard Source: @givisok, 2026-10-26 Source: @jjb01957, 2026-10-27; server fans this strong need a high-current controller Source: @port00, 2026-10-18 |
| Arctic P12 Pro | 120 x 25 mm | 600 to 3,000 rpm | 6.9 mmH2O | 0.33 A | spec; 77 CFM. Fine one-per-pair at moderate caps; "quieter per watt" than a Noctua A12x25 Source: @wakeupmbrc1, 2026-10-09; not enough alone on a three-card shroud Source: @reala7096, 2026-10-08 |
| Arctic S8038-10K | 80 x 38 mm | 500 to 10,000 rpm | 51 mmH2O | not checked | spec. Used with V620 shrouds; loud Source: @unlikelypotato, 2026-10-06 |
| Supermicro FAN-0074L4 | 80 mm | not given | not given | not given | One per pair, quieter than typical server fans; rackmountnet listing Source: @c133, 2026-10-08 |
| Supermicro FAN-0163L4 | 80 mm class | not given | "about 8 times" the FAN-0074L4 | not given | 9 of them in a 1U server for training, about 85 dBA Source: @programmerboi, 2026-10-09 |
| Delta BFB1012EH | 97 x 94 x 33 mm blower (9733) | not checked | not checked | retail listings quote 2.94 A | Max 67 C on one card, not loud Source: @froggao, 2026-10-15 |
| Delta BFB1012VH | 9733 blower | not given | not given | 2.7 A | Moved less air through the card than two 40 mm AVC fans at 100 % PWM; only wins at low rpm, which is not enough for 200 W Source: @fafinet, 2026-10-16 |
| AVC DBTB0428B2G | 40 mm axial | not given | more than the BFB1012VH | about 1 A | Two per card; under 75 C; droning noise Source: @fafinet, 2026-10-13 |
| PLB08025B12HH, CF7525U12D | 80 mm and 75 mm blowers (the fans from Asus RTX 30 cards) | not given | not given | not given | The blowers the PluraliT housing is designed for; 60 C at 200 W at 50 to 60 % Source: @68x0, 2026-10-15 |
| San Ace 9BMB12P2G01 | 97 mm blower | high | not given | not given | Cools the card but extremely loud Source: @civiloid, 2026-10-13 |
| Phanteks T30 | 120 x 30 mm | not given | not given | not given | Recommended by @impactic Source: @impactic, 2026-10-06 and @djpandaman Source: @djpandaman, 2026-10-09 |
| Noctua A12x25 and 3,000 rpm Noctuas | 120 mm | not given | not given | not given | Contested @wakeupmbrc1: bad advice, Noctua SSO2 bearings have a low-frequency peak that is hard to muffle Source: @wakeupmbrc1, 2026-10-09; @legitsplit finds their noise profile more pleasant Source: @legitsplit, 2026-10-09; @djpandaman: an A12 can work at 180 W Source: @djpandaman, 2026-10-09. Two 3,000 rpm Noctuas leaked oil and went back under RMA Source: @froggao, 2026-10-26 |
Specs above for the Arctic fans are from Arctic's product pages, read 2026-10-30.
Printable ducts, shrouds and parts
Check the license before you re-share a file. Several parts below were posted only as Discord attachments with no license; ask the author before hosting them elsewhere.
| Part | Author | Where | License | Fits | Notes |
|---|---|---|---|---|---|
| 2x 120 mm fan to 4x CMP170HX/A100 PCIe intake plenum | Tim Kanarsky (@turbine) | Printables 1801425 | CC BY-SA (Printables) | four cards side by side; also A100 PCIe | Proven The most used shroud in the Discord Source: @djpandaman, 2026-10-08 Source: @forte9000, 2026-10-08 Source: @turbine, 2026-10-08. Author's notes: two Arctic S12038-4K are "the minimum spec"; 15 cm from card fronts to fans; 6x M3 6 mm bolts to the cards, 3x M3 8 mm bolts with nuts to the plenum, a 2.5 mm hex driver; the plenum blocks the power connector clips (use pliers); print in a structural profile; Onshape source linked from the page. |
| Horizontal, symmetrical remix of the plenum | @mountains1605 | MakerWorld 3409347 | not checked (a remix of a CC BY-SA model should carry CC BY-SA) | four cards, horizontal mounting | Source: @mountains1605, 2026-10-08 |
Three-card version for one 120 mm fan (170HX_v19_60mm_M2.zip) |
@reala7096 | Discord attachment | not stated | three cards; can sit side by side | Source: @reala7096, 2026-10-08 |
| cmp-170hx-duct: parametric push duct for one 120 mm fan over two cards (presets V30 for cards one slot apart, A30 for adjacent cards) plus exhaust collectors | @adrienbrault | GitHub with interactive previews | MIT; the CFD solvers folder is GPL-3.0 | two cards per fan; for four adjacent cards two ducts do not fit Source: @adrienbrault, 2026-10-20 | Designed for an Arctic P12 Pro PST CO. A30 is printed and in use by the author; V30 and the collectors were never printed, so their fit and gains are model results only. Print fan face down, no supports, PETG 0.28 mm, about 68 to 72 g; never scale the STL, change duct/params.py instead; STLs are Git LFS objects. The author's own CFD says a pulling exhaust collector adds about 35 to 40 % airflow, more than any duct shape change; no measured GPU temperatures published yet. Source: repo github.com/adrienbrault/cmp-170hx-duct by adrienbrault on GitHub, posted by @adrienbrault, 2026-10-20 (MIT; solvers GPL-3.0)
|
| 170hx-fanduct: OpenSCAD generator from one rectangular fan opening to N stacked card inlets | @dkpoulsen (thcleaner) | GitHub | none stated | default 3 cards at 40 mm pitch, 120 x 240 mm fan opening | Override from the command line: openscad -D gpu_count=4 -D gpu_pitch=50 -o duct.stl 3x170hx_duct.scad. No STL shipped. @c133 made a 120 mm fan to 3-card duct with it. Source: repo github.com/dkpoulsen/170hx-fanduct by dkpoulsen on GitHub, posted by @c133, 2026-10-08 (no license stated) Source: @c133, 2026-10-08
|
| Nvidia CMP170hx Integrated Fan housing (replaces the top cover, holds a blower) | PluraliT (@port00, RazK) | Thingiverse 7409424; updated version on Cults3D (pay what you want) | CC BY-NC per a Thingiverse mirror listing (check on Thingiverse); Cults3D license not checked | one card; slim version fits 5-GPU PCIe switch boards Source: @moliminous, 2026-10-10 | Proven 60 C at 200 W with a PLB08025B12HH blower at 50 to 60 % Source: @68x0, 2026-10-15; five cards at 33 to 37 C with the Cults3D version Source: @moliminous, 2026-10-10. For PLB08025B12HH or CF7525U12D blowers; tested on PETG; three parts, part 3 needs supports. Updated with screw sizes and more intake at the back for tight spaces Source: @port00, 2026-10-10. Recommended by @djpandaman and @68x0 Source: @djpandaman, 2026-10-15 Source: @68x0, 2026-10-15 |
| Nvidia A-Series Blower Fan Adapter | Level1Techs forum | Level1Techs thread | not checked | A-series cards | Linked by @domrockt2 Source: @domrockt2, 2026-10-15 |
shroud_one_piece_v6.stl, IO_shield_adapter_pants_vlogo.STEP, IO_shield_adapter_pants_little_guyo.STEP |
@backwardsminer | STL, STEP files | not stated | one card | Source: @backwardsminer, 2026-10-15 |
| High-efficiency 120 mm fan grill | Noctua | Printables 1096961 | CC BY-NC-SA (Printables) | 120 mm fans | Theory Untested on these cards; @dottoreargus wants to see if it lowers shroud noise Source: @dottoreargus, 2026-10-08 |
| AMD V620 80 mm fan shrouds | unknown | not linked | unknown | "almost" fit; heat gun or a resized STL | Source: @unlikelypotato, 2026-10-06 |
| Exhaust shroud that plugs into the PCIe slot between two cards | @froggao | photo only | none posted | two cards | Source: @froggao, 2026-10-26 |
| Sliding plenum for 2 to 4 cards (first used on a Radeon V620 build) | @alanoo | files promised, not yet posted | unknown | 2 to 4 cards; a dual-fan 4-card version was printing | Source: @alanoo, 2026-10-24 |
| LLM-generated stackable 9733 funnel | @kellertreppe123 | not posted | unknown | stacked cards | Latches came out on the wrong side; funnel split diagonally for throughput Source: @kellertreppe123, 2026-10-13 |
Push and pull plenums for an eight-card stack (frame_plenum_8x170hx_3fan_short_bottom.stl, _EPS_bottom_2.stl, _EPS_top_3.stl, _short_top.stl) |
@neapopa | Discord attachments | not stated | eight cards; the rear plenum needs the slot brackets removed | Designed with an LLM: one push and one pull plenum that slot tightly onto the cards with no air bleed and no screws, each split in two for small print beds. The first print needed thicker floating regions at the EPS hole; it still lacks a groove for the EPS clip, so it is a bit hard to seat (a heat gun helps), and the thin joints were heat-gunned together. "This works well for me, I guess because there is no room for air leakage." Fans: Delta AFB1212SHE in front, Bitmain L3+ fans for exhaust. Source: @neapopa, 2026-10-22 |
Mount for San Ace 9BMB12P2G01 blowers (cmp170hx_sanyo9bmb_adapter.zip) |
@civiloid | Discord attachment | not stated | one card per blower | An LLM modelled it from the vendor's STL files and the card outline; it took about five attempts to get something printable. The blowers cool the card, but @civiloid gave up on them for noise. Source: @civiloid, 2026-10-13 |
Plate that ties the backs of four cards into one block (frontplateminimal.stl) |
@mountains1605 | Discord attachment | not stated | four cards whose fronts share an intake shroud | Makes maintenance easier on an open, mining-rig style frame: without it the card backs wander. Source: @mountains1605, 2026-10-11 |
Front shroud for a 120 mm pull fan on an Octominer chassis (plenum_other_side_1x120_50mm_v11.stl) |
@reala7096 | Discord attachment | not stated | Octominer X12 Ultra | "Pretty basic", made for his own Octominer build. Source: @reala7096, 2026-10-15 |
- @squishysquare sells 3D-printed shrouds, fans and power cables as ready-to-go kits for two and four cards (CMP 170HX, A800, A100, A30), arranged by DM, and offers to set up US buyers. Source: @squishysquare, 2026-10-04 One buyer found the EPS cables of his PSU did not work with the kit and had to order different ones. Source: @ikyo, 2026-10-04 @con3po7016 bought 80 mm shrouds from @squishysquare's eBay account and they worked well; he suggests a 1.5 mm larger fan opening so the fan can be wrapped in rubber against vibration (one wrap of self-fusing tape stretched 2x was right). Source: @con3po7016, 2026-10-08
- Contested @squishysquare sells dual 40 mm shrouds rather than 120 mm ones because, he says, anything larger than 92 mm cannot push enough air through the cards Source: @squishysquare, 2026-10-16; many owners above cool with 120 mm fans on plenums Source: @mountains1605, 2026-10-08
Source: @wakeupmbrc1, 2026-10-16.
Sealing the air path
- Reported Taping over gridded slot brackets in a 2U server took 10 C off a card at 250 W. Source: @mlx4, 2026-10-20
- A hot printed shroud usually leaks at the card slot and around the power cable. Seal with electrical tape, aluminium ducting tape Source: @dottoreargus, 2026-10-07
Source: @analyticeth, 2026-10-07, or VHB and gaffer tape Source: @globalapathy, 2026-10-07. - Small gaps where the shroud meets the power connector did not need filling for @mountains1605 Source: @mountains1605, 2026-10-08.
- Contested Cards in a GPU server with empty slots: @zorg33 says fill every slot or print shrouds, since airflow efficiency is low without them Source: @zorg33, 2026-10-20; @markshowa says sealing the openings is enough, pressure rises and the air flows Source: @markshowa, 2026-10-20. Nobody posted a test.
- Four cards side by side make the middle two hottest; @cibernox leaves an empty slot in the middle Source: @cibernox, 2026-10-08. With several cards on one plenum, partly blocking the cooler cards evens out the airflow Source: @reala7096, 2026-10-08.
Fan power and control
- Reported The fan header footprint on the card is vestigial. @port00 populated it: PWM does not work, the fan runs at full speed, and the VBIOS has no fan curve, so it can never control a fan. Source: @port00, 2026-10-22 Even the A100 PCIe board has no PWM controller. Source: @endsjeremy, 2026-10-22
- Fan current: 4,000 rpm server fans (0.6 A max) run on a decent controller or a 3 A motherboard header; 8,000 rpm fans need a high-current controller. Source: @port00, 2026-10-18
Source: @froggao, 2026-10-19 Two channels of @alanoo's controller could not run S12038-4K fans at full speed. Source: @alanoo, 2026-10-26 - Proven A 10-port ARCTIC hub, SATA powered, with a USB 2.0 port that works on Linux, about $11, per-port PWM. Source: @68x0, 2026-10-16
Source: @globalapathy, 2026-10-16 Note: the link posted, Amazon.ca B0887VG14J, is ARCTIC's "Case Fan Hub" distributor, while USB and per-port PWM are features of ARCTIC's 10-port Fan Controller (ACFAN00351A); check which one you buy. @68x0's full parts list: a PLB08025B12HH-class blower (AliExpress 1005008497738205), a mini-GPU to full-size PWM pigtail (AliExpress 1005001374897318) and the hub, with a service that sets fan speed from GPU power and temperature. Source: @68x0, 2026-10-16 - ARCTIC 10-port Fan Controller ACFAN00351A: can follow the GPU temperature as its PWM curve Source: @globalapathy, 2026-10-22, but needs a Linux kernel update or patching for that Source: @globalapathy, 2026-10-09. It is limited to 2 A per port and 4.5 A in total, so only about two 2 A server fans Source: @globalapathy, 2026-10-22
Source: @port00, 2026-10-22. Feed fan power from elsewhere and route only PWM and tach through it Source: @port00, 2026-10-22, or use a heavy-duty controller such as the Coolboxx II (Tindie) Source: @port00, 2026-10-22. @port00's easy option is a cheap AliExpress fan controller. Source: @port00, 2026-10-22 - Other options: Y-cables and Molex or SATA power Source: @mlx4, 2026-10-15, an external PWM hub Source: @wakeupmbrc1, 2026-10-15, Aquacomputer Quadro or Octo for remote control Source: @wakeupmbrc1, 2026-10-15, two splitter cables on WRX80 headers (12 V 2 A each, so two 0.47 A fans per header is fine) Source: @.henrykrinkle, 2026-10-26.
- A BMC cannot read the temperature of a card on a generic passive riser (it shows as an add-in card, not a GPU), so it cannot drive the fan curve for that card; set that curve in software after the OS boots. Source: @fafinet, 2026-10-21 See Thermal limits.
Control strategy
- Regulate on the highest core or HBM temperature of all cards. Source: @alanoo, 2026-10-26 170tune's serving configuration requires the fan to be driven by HBM temperature, because a passive card heat-soaks and corrupts a memory overclock above about 75 C. Source: repo github.com/cachenetics/170tune by cachenetics on GitHub, posted by @rh3dai, 2026-10-04 (MIT license)
- Hold a thermal target instead of following a curve: keep every card's memory or core at 75 C regardless of load. It kept @con3po7016's server fans at their minimum. Source: @con3po7016, 2026-10-15
- Write a small service that sets fan speed from GPU power and temperature Source: @68x0, 2026-10-16; @feralzi wrote an app that controls a CMP fan through the motherboard and sets the power limit Source: @feralzi, 2026-10-17; @_orrorin keeps a small apartment quiet with 9733 blowers on a controller with custom curves Source: @_orrorin, 2026-10-20.
- sasha0552/nvidia-pstated (GPL-3.0) can call a script to switch external fans on when a card overheats or goes to a high performance state, and off after 15 minutes idle. Untested on the 170HX. Theory Source: repo github.com/sasha0552/nvidia-pstated by sasha0552 on GitHub, posted by @d3m3n7r, 2026-10-13 (GPL-3.0 license) See P-states and idle power.
- Reported Server BMC trap: if a fan script puts a Supermicro BMC into Full mode and writes the zone duty, the mode survives a power loss but the duty does not, so the next boot runs every fan at 100 % until the OS writes the duty again. Leave the BMC in its automatic (Optimal) mode while the OS is down, hand the fans back from a shutdown hook, and re-check the duty every 30 s. Below about 420 rpm the BMC flags a fan fault and holds 100 % until AC is removed, so never write a duty that low. Source: @darkmaniac7, 2026-10-27 (lab notes) PixelML documents the same Full-mode lesson. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 (Apache-2.0 license)
Noise
- A pre-shipping test video from the seller Alifond made @landontg worry the cards would be very loud; @djpandaman: the seller tests without caring about sound, and the right fan, shroud and the idle-power settings make a big difference. Source: @landontg, 2026-10-09
Source: @djpandaman, 2026-10-09 - Blowers: @civiloid's San Ace blowers at 70 % made a low-pitched noise that got through three closed doors and concrete walls; below 70 % they resonated and carried even further. He moved to water. Source: @civiloid, 2026-10-13
- 40 mm fans: a piercing drone at 30 to 40 % PWM Source: @fafinet, 2026-10-13; a server full of high-speed 40 mm fans is loud enough to injure. Source: @endsjeremy, 2026-10-16
- Server chassis fans have high floors: about 3,750 to 4,000 rpm on a Gigabyte G292 even with a fan script. Source: @voidgiraffe, 2026-10-15
Source: @voidgiraffe, 2026-10-15 - Theory Put the loud fan outside the box: an inline 6-inch duct fan (300 CFM class, for example this SUPROCKY inline fan) exhausting a dense server from elsewhere. Untested on these cards. Source: @roboalchemist., 2026-10-26
Source: @endsjeremy, 2026-10-26 - Theory Accept the noise and put a blower-cooled box in a closet at about 55 C Source: @kellertreppe123, 2026-10-08; Delta also makes 6,000 rpm fans at server noise levels Source: @lenxtroz, 2026-10-08.
- Theory Rewiring Arctic fans onto the stock shroud is said to be much quieter; nobody posted temperatures. Source: @danske1, 2026-10-26
- Reported Noctua industrial fans are loud under load. @munisix runs seven cards on a mining rack with riser cables and fan controllers; his Noctuas are very nice at idle, but nobody builds these rigs to sit idle, and under load they sound like a constant hair dryer. He finds his RTX PRO 6000 Max-Q blowers much quieter, calls cooling the CMPs "a massive pain", and would like to drop the risers and fit the cards in a case. Source: @munisix, 2026-10-25
Source: @munisix, 2026-10-25
Source: @munisix, 2026-10-25
Source: @munisix, 2026-10-25
Source: @munisix, 2026-10-25 - Theory @siriusrhi plans an eBay cooling shroud with one Noctua industrial 3,000 rpm fan per pair of cards (one fan per card on home-made brackets if a pair is too much), on motherboard fan control so the fans only spin up under load. Source: @siriusrhi, 2026-10-25
Source: @siriusrhi, 2026-10-25 @endsjeremy plans to make air-cooling systems for modded cards one of his first products. Source: @endsjeremy, 2026-10-24 - If you want silence, water-cool. Source: @djpandaman, 2026-10-09
Source: @lia_oc, 2026-10-26 See Water cooling.