Thermal limits
From CMP Wiki
More actions
The CMP 170HX runs hot for its power, and it has two temperatures that matter: the GPU core and the HBM memory. Under load the HBM is usually the hotter one, and it is the one that limits memory overclocks and corrupts data first. Keep the core under about 80 C and the HBM under about 80 to 85 C for daily use.
Thresholds
| Sensor | Temperature | What happens | Source |
|---|---|---|---|
| Core | about 76 C | throttling starts, according to a rough cooler model; keep core and HBM under 75 C | Theory Source: @zorg33, 2026-10-20 |
| Core | 80 C | PixelML's stop rule for core temperature | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 (Apache-2.0 license) |
| Core | above about 80 C | GA100 leakage rises with temperature and the card runs away thermally | Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 (no license stated) |
| Core | 85 C | maximum operating temperature (read from the card, 170tune); "the cards will throttle at 85" Source: @matstatman, 2026-10-09; "85 degrees is the maximum, so you're most likely bottlenecked by cooling" Source: @xeax, 2026-10-15 | Source: repo github.com/cachenetics/170tune by cachenetics on GitHub, posted by @rh3dai, 2026-10-04 (MIT license) |
| Core | 95 C | hardware slowdown Source: repo github.com/cachenetics/170tune by cachenetics on GitHub; the card falls off the bus and turns off Source: @analyticeth, 2026-10-07 | see cells |
| Core | 100 C | one card reached it by accident, cut its power as hard as it could, and still works | Source: @sojun80, 2026-10-06 |
| HBM | about 75 to 76 C | a memory overclock (NDIV 70 and up) starts to corrupt data; 170tune qualifies NDIV 70 only with HBM at or below about 76 C | Source: repo github.com/cachenetics/170tune by cachenetics on GitHub |
| HBM | 80 C | "hbm doesn't have issues until 80c or so" Source: @replacedriver47, 2026-10-10; "anything above 80 is too hot" Source: @spectakkular, 2026-10-09 | see cells |
| HBM | 85 C | PixelML's stop rule for memory; above 85 C JEDEC halves the refresh interval, so keep stock refresh settings | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub Source: repo github.com/vektorprime/170tune by vektorprime on GitHub, posted by @rh3dai, 2026-10-04 (MIT license; fork of cachenetics/170tune) |
| HBM | 95 C | maximum operating temperature Source: repo github.com/cachenetics/170tune by cachenetics on GitHub; "a few minutes" at 95 C corrupts data in the WPR Source: @bigptoughneigh, 2026-10-09 | see cells |
Where people disagree
Contested
- How hot is too hot: @mountains1605 called @arabus_74064's temperatures on Qwen3.8 Next Flash too hot, while @arabus_74064 thought them normal at full load. Source: @mountains1605, 2026-10-09
Source: @arabus_74064, 2026-10-09 @replacedriver47: the SMs are not the problem, the HBM is, and it is more prone to heat damage Source: @replacedriver47, 2026-10-09; @rrickk420: "you're frying your mem" Source: @rrickk420, 2026-10-09; @bigptoughneigh: memory is not always hotter than the core under load Source: @bigptoughneigh, 2026-10-09; @dottoreargus barely reaches 65 C on the same model Source: @dottoreargus, 2026-10-09. - @impactic thought the card shuts itself down above 95 C; @prefix10k thought 85 C is already bad. Source: @prefix10k, 2026-10-06
Source: @impactic, 2026-10-06 - The throttle point: 76 C by @zorg33's model Source: @zorg33, 2026-10-20 against 85 C by @matstatman Source: @matstatman, 2026-10-09 and the card's own maximum operating value Source: repo github.com/cachenetics/170tune by cachenetics on GitHub.
Without airflow the card dies fast
- Proven A card without a fan reaches 95 C quickly even at its 40 W idle, falls off the bus and turns off. Source: @analyticeth, 2026-10-07
Source: @nikhilr1696, 2026-10-07 @programmerboi's passive card stopped showing innvidia-smiat 85 C while idle Source: @programmerboi, 2026-10-16, and a new 10 GB card "clocked out in under 5 mins without cooling. way worse than V100" Source: @programmerboi, 2026-10-25. The heatsink is designed for high-airflow servers and does not work even at idle without it. Source: @civiloid, 2026-10-16
Source: @globalapathy, 2026-10-25 - Warning A failed fan means a dead card or worse; @nikhilr1696 only runs these cards supervised. Source: @nikhilr1696, 2026-10-07 Use a decent fan, not necessarily a blower. Source: @cibernox, 2026-10-07
Throttling behaviour
- When the cap or the temperature limit bites, the core clock drops and with it memory bandwidth and tokens per second. On a 10 GB card with all SMs unlocked, the power cap tanked memory bandwidth even with the clock locked. Source: @djpandaman, 2026-10-12 See Undervolting and power limits.
- Warning A 250 W cap did not bound a fully SM-unlocked 10 GB card. @djpandaman's card at 122 SMs briefly reported about 400 W of board power during benchmarks with the limit at 250 W (he was unsure whether the reading was measured or calculated), and all of it ends up in the same heatsink and VRM. @xeax worries the VRM MOSFETs get overloaded and doubts he could cool that load. On such a card, watch core and HBM temperatures under real load instead of trusting the cap, and undervolt (see Undervolting and power limits#10 GB cards and Power supplies and cabling#PSU sizing). Source: @xeax, 2026-10-23
Source: @djpandaman, 2026-10-23 - Prefill is where cards run hottest: @c133's cards spike over their cap during long prefills Source: @c133, 2026-10-09; PixelML's worst-airflow card reached 80 to 85 C core on long-context prefill at 180 W, which is why they lowered the cap Source: repo github.com/PixelML/club-170hx by PixelML on GitHub.
- Check whether the card is in hardware power-brake slowdown:
nvidia-smi -q | grep -A1 'HW Power Brake Slowdown'. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub - Reported A card cooled by the fan that came in the Alifond package reached 85 C at full fan speed running a 27B Q8 model on llama.cpp; @xeax: 85 C is the maximum, so it is cooling-bound. Source: @programmer_007, 2026-10-15
Source: @xeax, 2026-10-15 - Fixes, in order: improve the air path (see Fans and ducts), lower the cap (Source: @xeax, 2026-10-15:
sudo nvidia-smi -i 0 -pl 200), then consider water (Water cooling). These chips run 400 W on SXM4 boards; better cooling is what unlocks more of them. Source: @xeax, 2026-10-15
What temperatures people run at
| Cooling | Cap | Hottest reading | Source |
|---|---|---|---|
| Water block, full load | 250 W | 50 C | Source: @leenodie, 2026-10-26 |
| Water block, 360 mm radiator, fans at minimum | 180 W | 45 C | Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub |
| Water, general | not given | 50 to 60 C | Source: @endsjeremy, 2026-10-20 |
| Air, unlocked 64 GB card | 175 W | 60 C core, 65 C HBM | Source: @rrickk420, 2026-10-08 |
| One 80 mm fan per pair, no duct | 185 W | 63 C | Source: @c133, 2026-10-09 |
| S12038-4K push only, four cards, cold garage | 200 W | 62 C | Source: @alanoo, 2026-10-26 |
| Blower at 50 to 60 % | 200 W | 60 C | Source: @68x0, 2026-10-15 |
| P12 Pro on a shroud | 175 W | 65 C | Source: @rrickk420, 2026-10-07 |
| Sealed funnel | 150 W | 70 C | Source: @globalapathy, 2026-10-07 |
| S12038-4K push-pull, 24 C room | 300 W | HBM 72 C | Source: @wakeupmbrc1, 2026-10-16 |
| P12 Pro push and pull on 3-card plenums, 10 minutes full power | full | 80 C | Source: @reala7096, 2026-10-08 |
| One card, 6-way inference soak for 10 minutes | 200 W | HBM 84 to 85 C, GPU 78 C | Source: repo github.com/cachenetics/170tune by cachenetics on GitHub |
| Four cards, GLM-5.3-Flash TP4 | 165 W | HBM 82 to 83 C | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub |
| Eight cards in a server with a fan wall, same cap, by slot | 180 W | 58 to 68 C on two slots per CPU, 81 to 84 C on the other two | Source: @darkmaniac7, 2026-10-28 (lab notes) |
| Four idle cards, one 80 mm blower on a printed duct | idle | core 37 to 38 C, memory 41 to 51 C | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub |
| Sandwiched cards with shared intake | under load | 34 to 42 C | Source: @moliminous, 2026-10-11 |
Reading the sensors
nvidia-smi -q -d TEMPERATURESource: @xeax, 2026-10-15, or for loggingnvidia-smi --query-gpu=index,temperature.gpu,temperature.memory,power.draw --format=csvSource: repo github.com/PixelML/club-170hx by PixelML on GitHub. Watchtemperature.memory: HBM is usually the warmest sensor under load. Source: repo github.com/cachenetics/170tune by cachenetics on GitHub- The card has about 10 temperature sensors, including one per HBM stack. Source: @matstatman, 2026-10-05 matstatman/hbmmon shows per-stack temperature, but it writes to the HBM debug interface by default and needs a custom driver; use its read-only mode on a stock driver. Source: repo github.com/matstatman/hbmmon by matstatman on GitHub, posted by @matstatman, 2026-10-05 (no license file)
- The 10 GB card's memory runs a few degrees warmer than the 8 GB card's (two more half stacks active): 41 C against 37 C, both at about 33 W idle, after a repaste; before the repaste the gap was 13 C, which pointed at the thermal interface. The owner used PTM on the 8 GB card and then ran out; what went on the 10 GB card is not stated. Source: @ghostinthedatacenter_87281, 2026-10-13
Source: @colo2442, 2026-10-13
Source: @djpandaman, 2026-10-13
Source: @ghostinthedatacenter_87281, 2026-10-13
Source: @ghostinthedatacenter_87281, 2026-10-13 - A server BMC cannot read the temperature of a card on a generic passive riser, so it will not ramp the fans for it. Source: @fafinet, 2026-10-21
Room and inlet temperature
- Your intake air sets the floor. Ivan's eight-card server raises its power caps only while the inlet is below 24 C and lowers them above 27 C; at about 2 kW into a room, the room became the limit before the fans did. See Chassis builds. Source: @darkmaniac7, 2026-10-27 (lab notes)
- Twelve cards on dedicated circuits with an air conditioner pointed at them is still the warmest and loudest room in @gpmickey's house. Source: @gpmickey, 2026-10-19
- Warning Do not feed cold winter air or sub-dew-point coolant to the cards: condensation. Source: @alanoo, 2026-10-20
Source: @gnasherrr, 2026-10-26 - Paste and pads matter for the last few degrees; see Water cooling for opening the card.