Operations Cooling

Cooling the card

What this page covers

Every cooling solution that has been measured on a CMP 170HX (or on an A100, which shares the PCB and the heatsink), what wattage each one actually removes, which marketed claims are false, and a recommendation table by target power limit. The thermal limits, sensor list and temperature reference tables live on Thermals.

The headline, because it is the most expensive mistake people make:

Caution

The 3.24 W snail fan sold as an "A100 cooler" does not do 300 W

The widely sold 3D-printed "A100 cooling" adapter bundled with a 3.24 W, 12 V snail fan is advertised as blowing off 300 W. First-hand measurement by a purchaser puts it at 150-180 W maximum, at maximum duty cycle, powered directly from the PSU rather than throttled by a header. That is well short of the card's 250 W nominal envelope and nowhere near the 300 W OC-VBIOS envelope. Nobody contradicted the measurement. A second participant separately queried whether a radial fan can do this job at 3 W at all, noting that most radial fans of the required class draw at least 3 A (about 30 W).

The card is fully passive, reports Fan Speed : N/A, and will not protect itself with a fan it does not have. Sizing rule: pick your cooler for the power limit you intend to run, using the measured column below and not the marketing column.

---

Five rules that fall out of the measurements

1. Pull, do not push. Sleeve-style shrouds that push air through the card produce "considerable blow back". Pulling air through the fin stack works much better in every first-hand account.
2. Screw the duct on. Friction-fit printed ducts "simply fall off" in use, in both single- and dual-GPU versions. The card has screw holes; use them.
3. Wattage is the proxy, not RPM. The useful spec is how many watts a solution has been shown to remove while holding 70 C core / 75 C memory, not the fan's rated speed.
4. A single 40 mm fan fails, whatever its RPM. One Arctic S4028-15K gives a 90 C hotspot; two give 70 C GPU / 76 C hotspot. An 8000 RPM 40 mm fan was abandoned by its own prospective buyer once the 80 C throttle onset was discovered ("guess 8k rpm won't do").
5. Cool the VRM too. Any solution that only covers the die, including a deshroud plus a pressure-mounted tower cooler, still needs a separate 80-120 mm fan over the power stages.

---

Blowers, measured

Solution	Measured result	Conditions	Confidence
Printed "A100 cooling" adapter + 3.24 W 12 V snail fan	150-180 W maximum (advertised 300 W)	max duty, direct PSU feed	high
EFB0251S3 blower, also rated 3.24 W	saturates at 73 C at 200 W	-pl 200, sustained 100% load	medium
San Ace B97, 1.85 A (BFB1012VH) + USB fan controller, curve capped at 66%	below 65 C at 250 W sustained	curve driven from hotspot; it hunts, die temp is the better input	medium
San Ace B97, 1.5 A + pot-driven PWM controller (PCIe or SATA powered)	claimed 350 W, unmeasured on a 170HX	unusable at full blast without a controller	low
9733S 12 V 2-wire blower (97 x 97 x 33 mm) + Level1Techs printed shroud	idle 64 C to 36 C; load 80+ C to 40-50 C	on an A100, same PCB and cooler, LLM workloads	medium
AC Infinity S2-class blower, no shroud	150 W OK, throttles at ~180 W	on an A100 (85 C limit), large air-leak paths	medium
AC Infinity S2-class blower, no shroud	200 W stress test passed	on an MI100	medium
Cheap 40/60 mm blower taped on with aluminium tape	no temperature posted	reported surprisingly sturdy, and loud	low

Note

Open problem: two nominally identical 3.24 W blowers, two different outcomes

The printed adapter with a 3.24 W snail fan tops out at 150-180 W. An EFB0251S3, also stated at 3.24 W, holds a card at 73 C at a 200 W limit under sustained full load. Same nominal fan power, materially different result. Duct sealing and the static-pressure path are the obvious explanation, and the tester who measured the unshrouded case reported "There are huge spots for leaks", but nobody has run the two side by side on the same card at the same power limit. Until someone does, treat the 3.24 W class as a 150-200 W class and assume duct quality is doing most of the work.

Treat 250 W as the proven figure for the B97 and 350 W as an unverified claim. The 350 W number is confidently asserted and supported by the same blower cooling a 350 W FPGA for a year, but no measured 350 W log on a 170HX exists.

---

Axial fans and shrouds, measured

Solution	Measured result	Conditions	Confidence
1 x Arctic P12 Pro PST 120 mm through a printed shroud, feeding 2 cards	32 C idle at ~34 W each	1000 RPM (fan capable of 3000), electrical tape added around the shroud to raise static pressure	high
2 x Arctic P12 Pro PST CO 120 mm in printed ducts with internal splitter vanes, feeding 4 cards	under 65 C	-pl 160 per card, "not much noise"	medium
1 x 120 mm @ 3000 RPM, ducted	73 C peak	stock clocks, "a bit too loud"	high
Cooling unstated (one owner, one setup)	~70 C ±2 default clocks; 75-77 C overclocked	up to 300 W. The fan was never named in the source thread	high
Arctic 12038-4K 120 mm	~85 C, "cannot cool it any better without water"	memory-overclock gpu_burn; HBM errors appeared within the first couple of minutes, cause unresolved	medium
1 x 140 mm Noctua NF-A14 industrialPPC + repurposed P100 shroud, 3 cards	~38 C idle, under 80 C loaded	wattage-driven custom fan curve, quiet at idle	high
2 x stacked Noctua 140 mm	~80 C down to ~70 C	adding the second fan	high
80 mm 10k-RPM server fan @ 3500 RPM	under 70 C core / 75 C memory at 200 W	same setup handles 300 W without difficulty	medium
80 mm server fan @ 7500 RPM	70 C core / 75 C memory at full 300 W	with an overclock	high
80 mm server fan @ 35% speed	within 70/75 C	-pl 175	high
2 x Arctic S4028-15K 40 mm on a custom 2-slot bracket	never above 70 C GPU / 76 C hotspot	curve set to 100% at 70 C, fans hit 15000 RPM	high
1 x Arctic S4028-15K 40 mm	90 C hotspot, rejected	same bracket	high
Two-fan printed shroud shipped with some cards	never above 60 C	under half fan speed, card not yet unlocked, workload unspecified	medium
Stock passive heatsink in a datacenter chassis (Gigabyte G292-Z20, 80 mm fans)	peak 60 C at 254 W	8-card rental; "louder than a jet engine"	medium
Deshroud + 2-3 x 120 mm Noctua directly onto the opened fin stacks	untested	proposed alternative to ducts	low
Deshroud + pressure-mounted tower cooler	untested	"gravity or zip ties works", no clamping pressure needed; VRM still needs its own 80-120 mm fan	low

The 80 mm high-RPM server-fan rows are the strongest air results in the corpus: they meet the 70 C / 75 C design target at the full 300 W envelope, and at 200 W they do it at only 3500 RPM.

---

Shrouds and ducts

Design	Verdict
Level1Techs A-series blower adapter (l1 a100 blower.stl, 52.3 KB)	Best-evidenced printed part. Support-free print that bolts to the existing screw holes at the far end of the card, with an angled low section beside the power connector for finger access. Produced the 64 C to 36 C idle result on an A100 with a 9733S blower. Caveat: the card becomes noticeably longer. Filament, print settings, airflow direction and durability were asked in-thread and never answered.
Thingiverse 5532715 (single and dual 120 mm)	Usable; use the "fixed" variant. The dual version is too long for some rigs.
CMP_170HX_Fan_Shroud_Fixed.stl / CMP_170HX_Dual_Shroud_Fixed.stl	Friction-fit, falls off over time, walls thin enough to be weak even in PETG. Only STLs exist, so modification needs a re-model rather than a STEP edit.
Cults3D cmp-170hx-fanduct (single and dual GPU)	Mechanical failure. Both versions "simply fall off".
LTT shroud	Rejected by users. "The ltt shroud is awful", "Lots of back-pressure."
Sleeve-style shrouds / push-through airflow	Rejected. Considerable blowback; will not work in 2-slot spacing. Flush bolt-on brackets will.
A5000/A6000 blower shrouds as a drop-in	Disputed, leaning against. The proposal rested on shared Ampere workstation PCB dimensions; the rebuttal cited different screw holes and no fan connector. Neither side posted a photo of an attempted fit.
A100 shrouds	Unresolved. Asserted to fit ("Only difference is the text right?"), never photographed or measured on a 170HX. Note that shared-PCB reasoning has already failed once here, with the A5000/A6000 claim.
Boring out the stock two-slot shroud for an integral fan	Works for a single card or spaced-out cards; fails for adjacent installs because two-slot cards leave no intake gap. Never built.

Practical gotchas reported repeatedly: several shrouds block access to the card's EPS power plug, so test-fit with the cable in place; third-party print quality is unreliable (one part described as 120 mm arrived as 140 mm); and use aluminium tape, not duct tape, if you are taping a blower on.

Note

Open problem: print material

PLA is reported as fine in practice by a multi-card user ("pla has been fine for all mine"), on the argument that the duct sits in cool intake air away from the heat source. Others recommend PETG or ASA as safer near GPU heat. Mildly disputed, no failure reported on either side.

---

Water cooling

Water is the only route that has produced sub-50 C load temperatures on this card.

Block	Measured result	Notes
Bykski N-TESLA-A100-X-V2 + 360 mm 3-fan radiator	30 C idle at 30 W; 45 C after 30 min at 180 W with fan and pump both at minimum speed	The only commercially available full-coverage block confirmed compatible. Also fits A100 40 GB PCIe, A30 24 GB and Tesla L40 (shared PCB). Uses standard G1/4 fittings. Lower temperatures are available at higher fan speed.
Budget generic full-cover block	48 C core / 58 C memory after a 1-hour stress test	8 GB card, in service about a year.
EK-PRO GPU WB RTX A100	no temperatures posted	Fits. Out of production.
Chinese V100 SXM2 flat-plate block	untested	Plain flat plate with no conforming surfaces; the MOSFET and inductor gaps would have to be filled by hand.
Arctic Liquid Freezer II 420 AIO + copper adapter plate	FAILED before a game finished loading	Photographed pump-out: cracked paste with bare copper across much of the die contact area. Attributed to poor mounting pressure or an incompatible adapter.

Caution

Fitment traps that destroy cards or waste money

• N-TESLA-A100-80G-X-V2 does not fit. The A100 80 GB is a different, incompatible PCB. Only the 40 GB-family block is correct.
• The V2 (all-metal) revision is required. The earlier non-V2 uses transparent acrylic.
• SXM waterblocks do not fit PCIe cards at all. At least a third of A100s sold were SXM. There are also two different PCIe A100 block designs, because the A100 80 GB has no integrated heat spreader while other variants do: "if that waterblock is made for the variant without IHS it wont work properly."
• The correct V2 block ships with no manual and a hex wrench of the wrong size. Have a metric hex set ready.

Caution

The single most damaging waterblock installation mistake

Cover every unpopulated IC footprint within reach of the block's contact pillars with thermal pad before lowering the block. Because the 170HX is a depopulated A100 board it has far more bare footprints than a real A100, and the block's metal pillars can short across exposed copper pads and permanently kill the card. Pad the DrMOS MOSFETs left and right of the ASIC, the areas bottom-left and bottom-right of the ASIC, the PMIC to the right of the die between an inductor and a capacitor, and the two PMICs to the left of the die below the 3.3 µH inductor. Thermal paste (pea-sized) goes only on the GPU/HBM copper spreader. The author of this procedure lost a card about an hour after installation and this is their leading suspicion, competing with a pre-existing mining-wear fault; it was never definitively isolated.

Full teardown order, washer and bracket reuse, and the 15-minute pressurised leak test are on Physical mods.

Note

Open problem: HBM temperature on water

The one sweep that tried to lower HBM temperature on air concluded "getting lower hbm temps on air seems impossible". A request for equivalent waterblock memory-temperature data went unanswered, so nobody knows whether the HBM floor is a cooling limit or intrinsic to the package. The budget-block owner who measured 58 C memory already has the hardware to answer it.

---

Fan control and noise

There is no usable fan control on the card itself. Everything is external.

• No confirmed on-card fan header. An unpopulated pad that looks like a fan location was spotted on the card. Asked whether it is a 4-pin, one participant answered that it carries 12 V and ground only; a skeptic in the same thread said it "doesn't look like a fan con". Nothing was ever probed, and no photo-confirmed pinout and no working fan install was ever posted. Even if it is a header, the reported pins are power only, which would mean a fixed full-speed fan with no tachometer.
• Control the fan from the host. A USB or pot-driven PWM controller is what testers actually used. Drive the curve from die temperature, not hotspot, because hotspot-driven curves hunt.
• A wattage-driven curve was reported as the quietest approach on a 3-card rig with 140 mm industrial fans.

Noise, in the order testers described it:

Setup	Noise
1 x P12 Pro 120 mm at 1000 RPM feeding 2 cards	quiet
2 x 120 mm feeding 4 cards at -pl 160	"not much noise"
140 mm industrialPPC on a wattage-driven curve	quiet at idle
1 x 120 mm at 3000 RPM ducted	"a bit too loud"
San Ace B97 at full blast without a controller	unusable
2 x S4028-15K 40 mm at 15000 RPM	loud by design
Stock passive heatsink in a datacenter chassis	"louder than a jet engine"

One more practical annoyance: a fixed-speed fan on a card with no fan control will "kick in at 100% speed during a system reboot", which is startling the first time.

---

Recommendation by target wattage

Pick the row for the power limit you will actually run, then verify with an integer or memory benchmark, never a conventional FP32 burn-in (see Power and PSU for why FP32 only draws ~60 W on this card).

Target	Recommended	Evidence	Avoid
Idle / light (~30-45 W)	Any 120 mm fan through a screw-on duct at low RPM	1 x P12 Pro at 1000 RPM held 32 C across 2 cards	Nothing at all: passive with no airflow leads to leakage runaway
150 W (-pl 150)	3.24 W-class blower with a well-sealed printed adapter, or a single ducted 120 mm	Snail-fan adapter proven to 150-180 W; AC Infinity S2 class OK at 150 W on an A100	Any unshrouded blower; leaks destroy the static-pressure path
160-200 W (-pl 160 to -pl 200)	1 x 120 mm per 2 cards in a vaned duct, or an 80 mm 10k server fan at ~3500 RPM	2 x 120 mm held 4 cards under 65 C at -pl 160; 80 mm at 3500 RPM held under 70/75 C at -pl 200; EFB0251S3 saturated at 73 C at 200 W	Single 40 mm fan (90 C hotspot)
250 W (stock cap)	San Ace B97 1.85 A + PWM controller, or 2 x 40 mm S4028-15K on a 2-slot bracket	B97 held below 65 C at 250 W sustained; 2 x S4028-15K held 70 C / 76 C hotspot; one owner sat at ~70 C ±2 at default clocks with the cooling never described	The 3.24 W snail-fan adapter (it tops out 70-100 W short)
300 W (OC VBIOS only)	80 mm server fan at 7500 RPM, or water	80 mm at 7500 RPM met 70 C / 75 C at full 300 W; one 30 min gpu_burn at a 300 W limit held 75-77 C with the cooling unstated (a single report, not reproduced)	Thin 120 mm fans, doubted for 300 W and especially for two cards. One Arctic 12038-4K owner sat at ~85 C and could go no further on air
Quietest at any wattage	360 mm loop with a compatible full-cover block	45 C at 180 W with fan and pump at minimum; 48 C core / 58 C memory on a budget block at load	AIO plus adapter plate: the one documented attempt pumped out and failed
Multi-card, dense	80 mm high-RPM server fans per card, or ducted 120 mm with splitter vanes	4 cards under 65 C on 2 x 120 mm at -pl 160; 8-card datacenter chassis peaked at 60 C at 254 W	Two single shrouds side by side: they will not fit two adjacent cards

Before buying anything, note that raising the power limit buys very little on this card (+2.8% BF16 from 250 W to 300 W, with temperatures already below 65 C), so cooling for -pl 160 to -pl 200 is a defensible choice on efficiency grounds alone. The measured efficiency peak of 1390 GFLOP/W was recorded at a 1400 MHz ceiling with a +350 MHz offset drawing about 134 W, but that cell was never validated with a memory sweep and sits between 1400/+325 (silent corruption) and 1400/+375 (fault), so it is not an operating point. Neither is the 1376 GFLOP/W cell at 1350 MHz / +300, which is also a single completed run that no pattern sweep ever gated. The highest sweep-validated efficiency point is the shipped eff profile, 1366 GFLOP/W at +250 / 1350 MHz. See Tuning.

---

Claims shown to be false

Claim	Status
The 3.24 W snail-fan "A100 cooling" adapter handles 300 W	False. Measured 150-180 W maximum.
Friction-fit printed ducts are adequate	False. Both single and dual versions fall off.
A single 40 mm fan is enough at high RPM	False. 90 C hotspot on an S4028-15K; the 8000 RPM proposal was abandoned by its proposer.
Sleeve shrouds pushing air through the card work	False. Considerable blowback; pulling works much better.
A5000/A6000 blower shrouds drop in	Contradicted (different screw holes, no fan connector), though not photographed either way.
The LTT shroud is a good option	Rejected by users for back-pressure.
90 C is a fine operating temperature	False. Throttle onset observed at 80 C; design target is 70 C core / 75-76 C memory.
Better cooling plus 300 W unlocks significant performance	False. Measured +2.8% BF16, with core and memory both under 65 C.
An AIO cold plate on the IHS via a copper adapter works	Failed in the one documented attempt: photographed pump-out.
SXM waterblocks fit with straps and a thick pad	False. Physically incompatible with PCIe cards.

---

See also

• Thermals: limits, sensors, and the full measured-temperature reference.
• Power and PSU: power limiting, and why FP32 burn-ins do not load this card.
• Power delivery: where the VRM sits and why it needs air.
• Physical mods: teardown sequence, waterblock install, capacitor mod.
• Tuning: the clock and power sweep that produced the efficiency numbers.