Cooling and power - Revision history

ImportBot: Cooling and power overview, from the cmpunlocker Discord (cooling-power items, 2026-10-04 to 10-28), cachenetics/170tune, PixelML/club-170hx, amoghmunikote/170th-Street and Ivan's lab notes; sources: @analyticeth, @nikhilr1696, @programmerboi, @rh3dai, @mlx4, @wakeupmbrc1, @zorg33, @darkmaniac7, @c133, @68x0, @_arcanum_, @artistosu (+5 more); GitHub PixelML/club-170hx, cachenetics/170tune

2026-10-30T08:51:08Z

Cooling and power overview, from the cmpunlocker Discord (cooling-power items, 2026-10-04 to 10-28), cachenetics/170tune, PixelML/club-170hx, amoghmunikote/170th-Street and Ivan's lab notes; sources: @analyticeth, @nikhilr1696, @programmerboi, @rh3dai, @mlx4, @wakeupmbrc1, @zorg33, @darkmaniac7, @c133, @68x0, @_arcanum_, @artistosu (+5 more); GitHub PixelML/club-170hx, cachenetics/170tune

← Older revision		Revision as of 08:51, 30 October 2026
Line 6:		Line 6:
	* {{Status\|Warning}} Never run the card without forced airflow, not even at idle. An idle card draws about 40 W and with no fan it reaches 95 C quickly; at 95 C it falls off the bus and turns off. {{Source\|handle=analyticeth\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1528513735575081070/1557491900934791409}}<br />{{Source\|handle=nikhilr1696\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1528513735575081070/1557492862344630364}} One card on passive cooling stopped showing in <code><nowiki>nvidia-smi</nowiki></code> at 85 C while idle {{Source\|handle=programmerboi\|date=2026-10-16\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1560492431768813611}}; another clocked out in under 5 minutes with no cooling. {{Source\|handle=programmerboi\|date=2026-10-25\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1563704621388529725}} See [[Thermal limits]].		* {{Status\|Warning}} Never run the card without forced airflow, not even at idle. An idle card draws about 40 W and with no fan it reaches 95 C quickly; at 95 C it falls off the bus and turns off. {{Source\|handle=analyticeth\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1528513735575081070/1557491900934791409}}<br />{{Source\|handle=nikhilr1696\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1528513735575081070/1557492862344630364}} One card on passive cooling stopped showing in <code><nowiki>nvidia-smi</nowiki></code> at 85 C while idle {{Source\|handle=programmerboi\|date=2026-10-16\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1560492431768813611}}; another clocked out in under 5 minutes with no cooling. {{Source\|handle=programmerboi\|date=2026-10-25\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1563704621388529725}} See [[Thermal limits]].
	* The power input is one EPS12V socket (the CPU 8-pin, 4x 12 V and 4x ground), not a PCIe 8-pin. Feed it with one EPS cable, or with both inputs of the bundled 2x PCIe 8-pin adapter on two separate PSU cables. Never use pigtails or one cable for both inputs. See [[Power supplies and cabling]] and [[Power and cabling warnings]].		* The power input is one EPS12V socket (the CPU 8-pin, 4x 12 V and 4x ground), not a PCIe 8-pin. Feed it with one EPS cable, or with both inputs of the bundled 2x PCIe 8-pin adapter on two separate PSU cables. Never use pigtails or one cable for both inputs. See [[Power supplies and cabling]] and [[Power and cabling warnings]].
	* Watch the memory (HBM) temperature, not only the core. HBM is usually the hottest sensor under load. ([https://github.com/cachenetics/170tune/blob/main/docs/power-cap-curve.md power-cap notes]) {{Source\|handle=rh3dai\|date=2026-10-04\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1556287917775655053\|repo=https://github.com/cachenetics/170tune}} ~~(MIT)~~		* Watch the memory (HBM) temperature, not only the core. HBM is usually the hottest sensor under load. ([https://github.com/cachenetics/170tune/blob/main/docs/power-cap-curve.md power-cap notes]) {{Source\|handle=rh3dai\|date=2026-10-04\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1556287917775655053\|repo=https://github.com/cachenetics/170tune\|note=MIT license}}

	== What airflow it needs ==		== What airflow it needs ==
Line 34:		Line 34:

	== Power facts ==		== Power facts ==
	* Stock power limit is 250 W on the common 8 GB VBIOS (92.00.67.00.01); some 8 GB cards ship with a 300 W VBIOS (92.00.6D.00.0A). Both share the board part 900-11001-0108-000. {{Source\|handle=c133\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562161732137386127\|repo=https://github.com/PixelML/club-170hx~~}} (~~Apache-2.0)		* Stock power limit is 250 W on the common 8 GB VBIOS (92.00.67.00.01); some 8 GB cards ship with a 300 W VBIOS (92.00.6D.00.0A). Both share the board part 900-11001-0108-000. {{Source\|handle=c133\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562161732137386127\|repo=https://github.com/PixelML/club-170hx\|note=Apache-2.0 license}}
	* Most people serve at a cap between 140 W and 200 W. A four-card GLM-5.3-Flash TP4 sweep found 140 W only about 1 % slower than 150 W for 8 % less power ~~<span class~~=~~"cmpw-source">Source: [~~https://github.com/PixelML/club-170hx ~~PixelML/club-170hx]</span>~~; 175 W is the knee in the 170tune single-card curve ~~<span class~~=~~"cmpw-source">Source: [~~https://github.com/cachenetics/170tune ~~cachenetics/170tune]</span>~~. See [[Undervolting and power limits]].		* Most people serve at a cap between 140 W and 200 W. A four-card GLM-5.3-Flash TP4 sweep found 140 W only about 1 % slower than 150 W for 8 % less power {{Source\|repo=https://github.com/PixelML/club-170hx}}; 175 W is the knee in the 170tune single-card curve {{Source\|repo=https://github.com/cachenetics/170tune}}. See [[Undervolting and power limits]].
	* Short spikes go above the cap: up to 380 W on cards capped at 250 W. {{Source\|handle=n2dcrypt\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1556771516803321877}} Size PSUs for that.		* Short spikes go above the cap: up to 380 W on cards capped at 250 W. {{Source\|handle=n2dcrypt\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1556771516803321877}} Size PSUs for that.
	* Idle draw is 30 to 40 W per card stock; patched idle settings reach 23 to 26 W. See [[P-states and idle power]].		* Idle draw is 30 to 40 W per card stock; patched idle settings reach 23 to 26 W. See [[P-states and idle power]].
	* Undervolting (shifting the V/F curve) cut power 24 % at a fixed 1,350 MHz on the 170tune reference card. It needs a VBIOS that exposes the VF offset; the 250 W 8 GB VBIOS and the 10 GB VBIOS do not. ~~<span class~~=~~"cmpw-source">Source: [~~https://github.com/cachenetics/170tune ~~cachenetics/170tune]</span>~~		* Undervolting (shifting the V/F curve) cut power 24 % at a fixed 1,350 MHz on the 170tune reference card. It needs a VBIOS that exposes the VF offset; the 250 W 8 GB VBIOS and the 10 GB VBIOS do not. {{Source\|repo=https://github.com/cachenetics/170tune}}

	== Pages in this section ==		== Pages in this section ==

ImportBot: Cooling and power overview, from the cmpunlocker Discord (cooling-power items, 2026-10-04 to 10-28), cachenetics/170tune, PixelML/club-170hx, amoghmunikote/170th-Street and Ivan's lab notes; sources: @analyticeth, @nikhilr1696, @programmerboi, @rh3dai, @mlx4, @wakeupmbrc1, @zorg33, @darkmaniac7, @c133, @68x0, @_arcanum_, @artistosu (+5 more); GitHub PixelML/club-170hx, cachenetics/170tune

2026-10-30T08:04:06Z

← Older revision		Revision as of 08:04, 30 October 2026
Line 1:		Line 1:
	~~''This section~~ is ~~being written from the Discord history~~.'' It ~~will cover fans, ducts~~, water ~~blocks, undervolting, p-states, power cables and PSUs~~.		The CMP 170HX is a passive data-center card. It has a large copper heatsink and no fan. It was built for a server that
			forces air through its fins, so on a desk or in a normal case it overheats unless you add that airflow yourself or fit a
			water block. This page is the short version. The detail lives on the pages listed at the end.

	~~Until then~~, see [[~~Start here~~]] and [[~~Contributing~~]].		== Before you power a card on ==
			* {{Status\|Warning}} Never run the card without forced airflow, not even at idle. An idle card draws about 40 W and with no fan it reaches 95 C quickly; at 95 C it falls off the bus and turns off. {{Source\|handle=analyticeth\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1528513735575081070/1557491900934791409}}<br />{{Source\|handle=nikhilr1696\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1528513735575081070/1557492862344630364}} One card on passive cooling stopped showing in <code><nowiki>nvidia-smi</nowiki></code> at 85 C while idle {{Source\|handle=programmerboi\|date=2026-10-16\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1560492431768813611}}; another clocked out in under 5 minutes with no cooling. {{Source\|handle=programmerboi\|date=2026-10-25\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1563704621388529725}} See [[Thermal limits]].
			* The power input is one EPS12V socket (the CPU 8-pin, 4x 12 V and 4x ground), not a PCIe 8-pin. Feed it with one EPS cable, or with both inputs of the bundled 2x PCIe 8-pin adapter on two separate PSU cables. Never use pigtails or one cable for both inputs. See [[Power supplies and cabling]] and [[Power and cabling warnings]].
			* Watch the memory (HBM) temperature, not only the core. HBM is usually the hottest sensor under load. ([https://github.com/cachenetics/170tune/blob/main/docs/power-cap-curve.md power-cap notes]) {{Source\|handle=rh3dai\|date=2026-10-04\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1556287917775655053\|repo=https://github.com/cachenetics/170tune}} (MIT)

			== What airflow it needs ==
			The fins are dense and the heatsink is long, so the fan has to push air through a restriction. What matters is static
			pressure and a sealed path, not open-air flow.
			* Air has to go through the fins, from the end of the card out through the slot bracket. Gaps where air can leak around the card cost you most of the effect. Taping over gridded slot brackets in a 2U server took 10 C off one card at 250 W. {{Source\|handle=mlx4\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1562008624908214304}}
			* Push-pull (a fan on each end of the card) helps a lot and lets you run fans slower. {{Source\|handle=wakeupmbrc1\|date=2026-10-16\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1560656839547424841}}
			* Airflow needed rises steeply with power. A rough model by @zorg33 put a 250 W cap at 2 to 3 times the airflow of 200 W. {{Status\|Theory}} {{Source\|handle=zorg33\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562110134271090689}}
			* The room matters too. Ivan's eight-card server put about 2 kW into the room; raising its fan wall from 10 % to 25 % did not cool the hottest cards (they stayed at 84 to 85 C) while the inlet air rose from 24 C to 26 C. {{Status\|Reported}} {{Source\|handle=darkmaniac7\|date=2026-10-27\|note=lab notes}}. See [[Chassis builds]].

			== Cooling options, cheapest to best ==
			{\| class="wikitable"
			! Option !! Rough cost !! Noise !! What people measured !! More
			\|-
			\| One 120 mm or 80 mm fan per pair of cards, taped or clipped to the card ends, no duct \|\| one fan per two cards \|\| low to medium \|\| under 63 C at a 185 W cap with Supermicro FAN-0074L4 80 mm fans {{Source\|handle=c133\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1528513735575081070/1558160764370690100}} \|\| [[Fans and ducts]]
			\|-
			\| 3D-printed intake plenum with 120 mm server fans (Arctic S12038-4K), push or push-pull \|\| filament plus two fans per four cards \|\| medium \|\| 300 W per card at 3,500 rpm with HBM at 72 C in a 24 C room, push-pull {{Source\|handle=wakeupmbrc1\|date=2026-10-16\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1560655488331546715}} \|\| [[Fans and ducts]]
			\|-
			\| Printed housing with a blower per card (9733 or 8025 blower) \|\| one blower per card \|\| loud at full speed \|\| 60 C at a 200 W cap with the blower at 50 to 60 % {{Source\|handle=68x0\|date=2026-10-15\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1560301937675599944}} \|\| [[Fans and ducts]]
			\|-
			\| Server chassis with its own fan wall \|\| a used GPU server \|\| loud; the BMC fan curve follows GPU temperature \|\| see [[Chassis builds]] \|\| [[Chassis builds]]
			\|-
			\| Dual-card AIO water cooler (two blocks, one radiator) \|\| about $250 to $400 per two cards before shipping (October 2026) {{Source\|handle=_arcanum_\|date=2026-10-04\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1556176261703925861}}<br />{{Source\|handle=artistosu\|date=2026-10-04\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1556178792190320735}} \|\| quiet \|\| 4 cards "cool and quiet" {{Source\|handle=civiloid\|date=2026-10-16\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1560654690897629264}} \|\| [[Water cooling]]
			\|-
			\| Full-cover water block on a custom loop \|\| about $140 to $220 per block plus the loop {{Source\|handle=peter643\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1545982693903306813/1562011865465487431}}<br />{{Source\|handle=_syn_thetic_\|date=2026-10-25\|link=https://discord.com/channels/1518634379604136178/1545982693903306813/1563734136701722645}} \|\| quiet \|\| 50 C max at 250 W full load {{Source\|handle=leenodie\|date=2026-10-26\|link=https://discord.com/channels/1518634379604136178/1545982693903306813/1564340979065163910}} \|\| [[Water cooling]]
			\|}

			== Power facts ==
			* Stock power limit is 250 W on the common 8 GB VBIOS (92.00.67.00.01); some 8 GB cards ship with a 300 W VBIOS (92.00.6D.00.0A). Both share the board part 900-11001-0108-000. {{Source\|handle=c133\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562161732137386127\|repo=https://github.com/PixelML/club-170hx}} (Apache-2.0)
			* Most people serve at a cap between 140 W and 200 W. A four-card GLM-5.3-Flash TP4 sweep found 140 W only about 1 % slower than 150 W for 8 % less power <span class="cmpw-source">Source: [https://github.com/PixelML/club-170hx PixelML/club-170hx]</span>; 175 W is the knee in the 170tune single-card curve <span class="cmpw-source">Source: [https://github.com/cachenetics/170tune cachenetics/170tune]</span>. See [[Undervolting and power limits]].
			* Short spikes go above the cap: up to 380 W on cards capped at 250 W. {{Source\|handle=n2dcrypt\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1556771516803321877}} Size PSUs for that.
			* Idle draw is 30 to 40 W per card stock; patched idle settings reach 23 to 26 W. See [[P-states and idle power]].
			* Undervolting (shifting the V/F curve) cut power 24 % at a fixed 1,350 MHz on the 170tune reference card. It needs a VBIOS that exposes the VF offset; the 250 W 8 GB VBIOS and the 10 GB VBIOS do not. <span class="cmpw-source">Source: [https://github.com/cachenetics/170tune cachenetics/170tune]</span>

			== Pages in this section ==
			* [[Fans and ducts]]: fan models, measured temperatures, printable shrouds and ducts, fan control.
			* [[Water cooling]]: blocks that fit, AIOs, installation, pads and paste.
			* [[Undervolting and power limits]]: V/F offset, power caps, measured watts against tok/s.
			* [[P-states and idle power]]: idle draw and how to lower it.
			* [[Power supplies and cabling]]: the EPS input, cable rules, PSU sizing per card count.
			* [[Thermal limits]]: core and HBM thresholds, throttling, what people run at.
			* [[Chassis builds]]: multi-card builds, fan walls and power distribution.

			== See also ==
			[[Troubleshooting]] · [[Power and cabling warnings]] · [[Platforms]] · [[Recipes]] · [[Benchmarks]]

			{{ImportFooter}}
			[[Category:Cooling and power]]

			[[Category:8 GB]]
			[[Category:10 GB]]

ImportBot: Wiki lane: site structure (templates, forms, help pages)

2026-10-30T07:00:53Z

Wiki lane: site structure (templates, forms, help pages)

New page

''This section is being written from the Discord history.'' It will cover fans, ducts, water blocks, undervolting, p-states, power cables and PSUs.

Until then, see [[Start here]] and [[Contributing]].