Hardware Thermals

Thermal design

What this page covers

The CMP 170HX's cooler, the temperatures the silicon will tolerate, the sensors the driver exposes, and every measured temperature in the corpus. This is the reference page: for choosing an actual fan or waterblock, see Cooling. For the power side of the same problem, see Power and PSU and Power delivery.

Two facts frame everything else. First, the cooler is fully passive: a bare heatsink with no fan and no fan header the driver can see, designed for the forced air of a 1U/2U server chassis. nvidia-smi reports Fan Speed : N/A in every capture ever posted, so the card will never spin anything up on its own and will happily cook itself in a quiet desktop case. Second, the card is hard to load: a conventional FP32 burn-in draws only about 60 W on a 250 W part, so a cooler validated that way is not validated at all.

The unlock is irrelevant here. A keyword sweep of the shipping unlocker and all 12 unreleased branches for thermal, pstate, power_limit, vid_pwm, clkdomain, freqDelta, 0x20340, MHz and watt returns nothing in any patch, script or config. Every thermal characteristic below is a property of the stock board and stock VBIOS.

---

The physical cooler

Property	Value	Basis
Cooler type	Fully passive bare heatsink, no fan, no shroud fan	direct observation
Slot width	Dual-slot, approximately 40 mm / 1.57 in thick	measured by owners fitting brackets, medium confidence
PCB length	27 cm bare; 29 cm including I/O shield and a plugged-in EPS connector	direct measurement
Clearance above the PCIe edge connector	roughly 5-6 mm (1/4 in)	measured, medium confidence
Heatsink bolt pattern around the die	57 x 68 mm centre-to-centre (68 mm vertical, 57 mm horizontal)	direct measurement
Die package	approximately 55 x 55 mm; die itself 826 mm²	measurement plus die database
Fasteners	Torx T10 and T15	teardown
Thermal interface, main areas	1.5 mm soft pad	one owner with the card open and a photo; a second guessed 2 mm without confidence
Thermal interface, inductors	3 mm soft pad	same source
Thermal interface, die	liquid thermal compound / liquid thermal pad	same source
Published CFM, static pressure, dBA, fin pitch	none exist	no vendor or teardown figure in any source

The 40 mm slot width matters for bracket shopping: A100/H100-style 40 mm blower brackets advertised for A100, H100, A20, A30, A40 and CMP 170HX measure 1.57 in wide and therefore fit standard 2-slot spacing.

!!! question "Open problem: nobody has measured this heatsink" No CFM, static-pressure, dBA or fin-pitch figure for the stock 170HX heatsink is published anywhere in this corpus. Every airflow number in this wiki is a community measurement of a fan-plus-card system, not a heatsink specification. The closest thing to an anchor is the two-120 mm-fans-for-four-cards result at -pl 160. Closing this needs one person with a known fan curve correlating fan operating point against steady-state die temperature at a fixed power limit.

---

Thermal limits

The first four come straight out of the driver on an unlocked 10 GB card (driver 610.43.02, captured 2026-07-25 at 18:32, P0, 18% utilisation). The last two rows are community observations, not driver-reported.

Limit	Value	Layer
GPU Shutdown Temp	98 C	hardware, driver-reported
GPU Slowdown Temp	95 C	hardware slowdown, driver-reported
GPU Max Operating Temp	85 C	driver-reported operating ceiling
Memory Max Operating Temp	95 C	driver-reported, HBM
Reported throttle onset	~80 C	a single unsourced reading, no telemetry posted, low confidence
GA100 memory throttle (recalled)	~85 C, "not on all bios"	two hedged recollections, low confidence

At that capture the card read GPU 34 C, memory 48 C, Fan Speed : N/A, and every clocks event reason Not Active, including HW Thermal Slowdown and HW Power Brake Slowdown, with every event counter at 0 us.

!!! question "Open problem: the throttle layering is inferred, not established" Four numbers are on the record for "when does it throttle": 80 C (one tester's own telemetry), 85 C (two GA100 recollections and the driver's Max Operating Temp), 95 C (driver Slowdown Temp) and "above 100 C" (described for a card whose cooling failed entirely). These may all be true at different layers: a soft VBIOS-level clock reduction at 80 C, a driver operating ceiling at 85 C, a hardware slowdown at 95 C, and runaway when there is no cooling loop at all. No one has posted a throttle log correlating clock reduction against temperature. What would settle it is a logged sweep through 75-95 C of nvidia-smi --query-gpu=temperature.gpu,clocks.sm,clocks_throttle_reasons.active --format=csv -l 1.

The community design target

Independently of the driver limits, testers converged on 70 C core and 75-76 C memory hotspot under load as the design target for a cooling solution. 90 C hotspot is treated as unacceptable and was the basis for rejecting a single 40 mm fan. In the exchange that settled this, a participant shopping for fans asked whether 90 C was acceptable for the chip, then reported reading that throttling starts at 80 C, and was answered "Absolutely not!". No telemetry was posted alongside the 80 C figure, so treat it as a reading rather than a measurement.

!!! note "Superseded" "90 C is a fine operating temperature for this chip" was the early consumer-GPU intuition. It is retired. Size cooling for 70 C core / 75 C memory, not for the 98 C shutdown point.

---

Sensors and telemetry

The card exposes GPU temperature and memory temperature, and nothing else thermal. There is no fan tachometer, no per-stack HBM sensor, no exposed VRM or hotspot-delta sensor, and no nvidia-settings fan control path because there is no fan.

# Static limits and the current reading
nvidia-smi -q -d TEMPERATURE

# Continuous log, one row per second, suitable for finding the real throttle point
nvidia-smi --query-gpu=timestamp,temperature.gpu,temperature.memory,clocks.sm,\
clocks.mem,power.draw,clocks_throttle_reasons.active --format=csv -l 1

Fields confirmed present on the 170HX:

Field	Confirmed value in a real capture
GPU Current Temp	34 C (idle, unlocked 10 GB)
GPU Shutdown Temp	98 C
GPU Slowdown Temp	95 C
GPU Max Operating Temp	85 C
Memory Current Temp	48 C (idle)
Memory Max Operating Temp	95 C
Fan Speed	N/A on every capture ever posted
Clocks Event Reasons	all Not Active at idle; all counters 0 us

nvtop reports the same sensors in one line, and is what most testers pasted:

GPU 1440MHz MEM 1890MHz TEMP 76C FAN N/A POW 278 / 300 W
PCIe GEN 1@ 4x
GPU 100%  MEM 57.534Gi/64.000Gi

A practical note from a tester building a fan curve: driving the curve from the reported hotspot sensor makes the controller hunt, and die temperature is the better control input.

---

Leakage feedback and thermal runaway

GA100 exhibits a genuine leakage-driven positive feedback loop. Higher junction temperature means higher CMOS leakage current, which means more heat, which means more leakage.

The cleanest measurement of it was taken by accident. One researcher dry-ran a card with a waterblock fitted but no coolant in it: idle draw started at about 40 W, climbed to 60 W at 80 C, and was still rising when the card was powered off.

!!! danger "Never power a card with a mounted-but-dry waterblock for more than 5 minutes" A dry block is worse than no block: it insulates. The measured 40 W to 60 W climb had not stabilised at 80 C. If you must dry-fit to check clearances, power off within 5 minutes.

Two consequences follow, one benign and one not:

• Benign: cooling the card better lowers its idle power, because it suppresses the leakage term as well as removing heat. This is the most likely explanation for the 30 W versus 44 W idle spread between testers.
• Not benign: if cooling fails outright, GA100 does not settle at its throttle point. One account from hands-on A100-class experience: "The thermal throttling runs at such a high power that if the cooling fails, it'll thermally throttle, but to a temperature above 100 degrees, that then consumes more power, thus increased temperature... and you end up with a volcano." The exact temperature is BIOS-dependent and was disputed in the same exchange, so treat the shape of the curve as established and the number as not.

---

Measured temperatures

Every temperature in the corpus, on a 170HX unless the row says otherwise. "Proxy card" rows are A100 or MI100 results included because the A100 40 GB shares the PCB and cooler.

Idle

Cooling	Temperature	Conditions	Confidence
360 mm radiator, fan and pump at minimum	30 C @ 30 W	Bykski N-TESLA-A100-X-V2	high
1 x Arctic P12 Pro PST 120 mm @ 1000 RPM, feeding 2 cards	32 C @ ~34 W each	printed shroud, taped for static pressure	high
Unstated (driver capture)	34 C GPU / 48 C memory	unlocked 10 GB, P0, 18% util	high
1 x 140 mm Noctua NF-A14 industrialPPC, 3 cards	~38 C	repurposed P100 shroud, wattage-driven curve	high
Unstated, cold room	29 C @ 37 W	locked 8 GB, driver 580.159.04	high
Unstated, 3 unlocked 40 GB cards	44 / 45 / 41 C	nvtop, 210 MHz core / 1215 MHz memory	high
Unstated, during MIG testing	61-62 C @ 44 W in P0	250 W cap	high
A100 proxy, 2 x 120 mm case fans at maximum	64 C	before shroud fitted	medium
A100 proxy, 9733S blower + printed shroud	36 C	after shroud fitted	medium

Under load

Cooling	Temperature	Load	Confidence
360 mm radiator, fan and pump at minimum	45 C after 30 min	180 W, FluidX3D FP32/FP16S, FMA disabled	high
Budget generic full-cover waterblock	48 C core / 58 C memory	1 h stress test, 8 GB card, in service about a year	medium
A100 proxy, 9733S blower + printed shroud	40-50 C	LLM load, down from 80+ C on 2 x 120 mm case fans	medium
Stock passive heatsink, datacenter chassis (80 mm fans)	60 C peak	254 W, 8-card rental	medium
Two-fan printed shroud shipped with a card	never above 60 C	under half fan speed, card not yet unlocked, workload never stated	medium
Unstated	61 C @ 208 W	sustained 100% load at the stock 250 W cap	high
San Ace B97 1.85 A + USB controller, curve capped at 66%	below 65 C	250 W sustained	medium
2 x Arctic P12 Pro PST CO 120 mm in ducts, feeding 4 cards	under 65 C	-pl 160 per card	medium
Large blower, 300 W A/B test	core and memory both below 65 C	BF16 at 300 W	high
80 mm 10k-RPM server fan @ 3500 RPM	under 70 C core / 75 C memory	-pl 200	medium
80 mm server fan @ 7500 RPM	70 C core / 75 C memory	full 300 W with an overclock	high
Cooling unstated	~70 C ±2	default clocks; same owner and same card as the 300 W row below	high
2 x Arctic S4028-15K 40 mm @ 15000 RPM, custom 2-slot bracket	never above 70 C GPU / 76 C hotspot	heavy load, curve 100% at 70 C	high
1 x 120 mm @ 3000 RPM, ducted	73 C peak	stock clocks	high
EFB0251S3 blower (3.24 W)	saturates at 73 C	-pl 200, sustained 100%	medium
Cooling unstated, 30 min gpu_burn	75 C rising to 77 C	unlocked 64 GB at a 300 W limit, 278/300 W drawn. One report, not reproduced	high
1 x 140 mm Noctua industrialPPC, 3 cards	under 80 C	LLM inference	high
2 x stacked Noctua 140 mm	~80 C down to ~70 C when the second fan was added	unstated load	high
Arctic 12038-4K 120 mm	~85 C	memory-overclock gpu_burn; owner reported no further headroom on air, HBM errors within the first couple of minutes	medium
1 x Arctic S4028-15K 40 mm	90 C hotspot, rejected	same bracket as the 2-fan row above	high
Unstated, SM-unlock GEMM ramp	62 C to 73 C in about 25-30 s	8 GB card, shows the ramp rate	high

The ramp figure is worth internalising: a 170HX goes from 62 C to 73 C in roughly 25 to 30 seconds once a real kernel starts. Idle temperature tells you almost nothing about whether your cooling is adequate.

The HBM floor on air

After a multi-configuration power and clock sweep on a card under a 140 mm shroud with a 120 mm fan at 3000 RPM, one tester concluded: "with the right tuning.. the temp baseline is right here.. you can save watts, but you wont be able to drop temps on air" and "getting lower hbm temps on air seems impossible". A 30-minute gpu_burn at those settings (EFF / Balanced +300MHz / 1400) completed with no errors.

!!! question "Open problem: is the HBM temperature floor a cooling limit or intrinsic?" A direct request for waterblock HBM temperature data went unanswered, so the floor is unbounded from below: nobody knows whether water gets HBM materially cooler than air. The single cheapest high-value experiment in this domain is for the owner of the budget waterblock (48 C core / 58 C memory over one hour) to re-run the same EFF / Balanced +300MHz / 1400 sweep with the memory sensor logged.

---

Thermals are usually not the limiter

This surprises people, and it changes how you should spend money on cooling.

• Raising the power limit from 250 W to 300 W on a well-cooled card gained +2.8% BF16 throughput (about 180 to 185 TFLOPS) with core and memory both below 65 C. The core simply does not want to clock higher.
• mmapeak on an 8 GB card with the OC VBIOS sat at 1470 MHz drawing only ~150 W with GPU-T reporting PerfCap: None, even with the power limit at 300 W. Neither power nor an exposed cap explains that ceiling, and it remains unexplained.
• The 80 GB configuration's instability is not thermal and not power-related: the failing cards never drew above about 80 W during the crashing workload.

!!! note "Superseded" "Delid the card and replace the die-to-IHS TIM" was proposed twice and closed by argument, not experiment: "We struggle to hit 250 W with GPCs disabled. That would have to be fixed first." Nobody has delidded a 170HX. The one reason to revisit it is diagnostic rather than thermal: if an HBM stack is physically too short to contact the IHS properly, that could be why it is disabled, which would matter most on the 8 GB card.

One genuine thermal fault mode does exist. Cards that have never been re-pasted or re-padded can throttle heavily: one owner with two 8 GB cards could only bench one, because the second "throttles a lot" and needed re-padding and re-pasting. If a card throttles far earlier than its neighbours under identical airflow, suspect the factory thermal interface before you suspect the silicon.

---

Airflow guidance

There is no measured requirement, only a set of rules that emerged from the measurements above. They are collected and compared against real coolers on Cooling.

Quantity	Value	Status
Target static pressure for full-power operation	20 mm	spec-derived recommendation, not measured
Arctic P12 Pro CO published spec	~7 mm static pressure, 130 m³/h	vendor spec; this fan nonetheless held 4 cards under 65 C at -pl 160
Minimum practical fan power	above 4 W, i.e. 0.35 A at 12 V	practical guidance, unchallenged, unmeasured
Preferred fan classes	radial high-static-pressure blowers, or 38 mm thick axial fans	thin 120 mm fans were doubted for 300 W and for two cards
Airflow direction	pull, do not push	consistent first-hand reports of blowback from push-through sleeve shrouds
Single 40 mm fan	insufficient at any RPM	1 x S4028-15K gives 90 C hotspot; 2 x gives 70 C / 76 C

Whatever cooler is used must also cool the VRM, not just the die. A deshroud plus a pressure-mounted tower cooler on the die still needs a separate 80-120 mm fan over the power delivery area. See Power delivery for where those parts sit.

---

See also

• Cooling: real coolers, measured, with a recommendation table by target wattage.
• Power and PSU: power limiting, idle draw, PSU sizing.
• Power delivery: the on-board VRM and rail topology.
• Physical mods: teardown, waterblock installation, and the capacitor mod.
• Tuning: the clock-ceiling and offset sweep behind the efficiency figures.
• Board and variants: SKU identification and board part numbers.