Building a Silent Rack for a Flat
Noise is the constraint that decides your homelab, and nobody budgets for it
Contents
The homelab hobby has an unspoken class divide, and it runs along the line between people with a basement and people with a flat. If you have a basement, noise is somebody else’s problem and you can buy the cheap 2U server with the 15,000 rpm fans. If your rack lives in the same 68 square metres as your bed, noise is the design constraint, and it quietly decides your hardware, your topology and how much of your hobby you are permitted to keep.
Mine lives in a flat. What follows is what actually moved the needle, roughly in order of how much benefit each change bought per euro, and a fair amount about why the obvious fixes don’t work.
The physics you can’t argue with
Three facts govern everything else.
Decibels are logarithmic, and your ear is worse than logarithmic. A 10 dB reduction is a tenfold drop in sound power and reads as roughly “half as loud” subjectively. So a change from 42 dB(A) to 38 dB(A) sounds like a marginal improvement on paper and is, in a quiet flat at 11 p.m., the difference between noticing the rack and forgetting it exists. Conversely, silencing one of two identical noise sources buys you 3 dB. If you have four loud fans and you fix one, you have achieved approximately nothing.
Airflow noise scales as roughly the fifth power of fan speed. This is the single most useful fact in the whole discipline. Halving fan rpm cuts the noise by around 15 dB while cutting airflow only in half. That trade is absurdly favourable, and it means the correct answer to almost every noise problem is more, larger, slower fans rather than better fans. Two 140 mm fans at 500 rpm move the air of one 120 mm fan at 1,400 rpm and are inaudible at two metres.
Small fans are physically incapable of being quiet at useful airflow. A 40 mm fan in a 1U chassis has to spin at 8,000–12,000 rpm to move anything, and at that speed the blade tips generate broadband noise plus a tonal whine at blade-pass frequency. No amount of foam fixes it, because the sound is generated at the blade and the frequency is high enough that it will find every gap in your enclosure. This is why 1U hardware and flats are fundamentally incompatible, and why the cheap eBay server that looks like such good value is a trap. I went through the same arithmetic from the power side in idle power draw, and the two constraints point at the same hardware.
Low frequencies go through walls; high frequencies bounce. Foam and soft furnishings absorb treble effectively and bass barely at all. Meanwhile mass and decoupling stop bass and do nothing for treble. Most people buy foam, hear the hiss disappear, and are then baffled that the low hum is still audible through a wall. They have solved the easy half.
Start by deleting the noise
Every euro spent on acoustic treatment is a euro spent working around a decision. Before you buy any foam, do the free things.
Get rid of the 1U and 2U kit. This is the big one and everything else is a rounding error next to it. A rack full of mini PCs or a single tower-format machine with 140 mm fans has a noise floor that treatment cannot reach on enterprise hardware. My rack lost about 18 dB(A) the day the 2U box left, which is more than every other intervention on this list combined.
Spin down what nobody is using. A rack that is on because it has always been on is a rack making noise for no reason. Wake-on-LAN on the machine that only runs a weekly job, and an automation that powers it down afterwards, turns an always-audible device into a silent one for 165 hours a week.
Move the spinning rust. Drives make two noises: a broad whoosh from the platters and, much worse, a clatter from head seeks. Seek noise is impulsive, which is the category of sound humans are best at noticing and least able to ignore. Consolidating the working set onto SSDs and leaving the array for bulk archive means the seeks happen twice a day instead of continuously.
Fix the fan curve before you touch the hardware. Most fan curves are configured by people optimising for a data centre, which means they ramp aggressively at temperatures your CPU finds relaxing. If your machine has a controllable PWM header, fancontrol from the lm-sensors package will do a decent job:
| |
The generated /etc/fancontrol is worth editing by hand. The key values:
| |
MINSTOP is the value below which the fan stalls; MINSTART is the higher value needed to get it moving from stationary. Both are per-fan physical properties and pwmconfig measures them for you. MAXPWM=200 deliberately caps the fan below its rated speed — the machine will thermal-throttle at 200 rather than scream at 255, and on a homelab box that throttles twice a year for ninety seconds, that is the correct trade.
Then verify it holds under load rather than trusting the config file:
| |
Measuring, so you know whether anything worked
Buy or borrow a sound level meter, or use a phone app with a calibration offset — the absolute number will be wrong by several dB but the differences are what you need, and differences are what a phone measures adequately. Take readings from the position that matters, which is the pillow, rather than 30 cm from the chassis where the vendor measured.
Take the room’s floor first, with everything off. A quiet flat at night sits around 28–32 dB(A); the fridge, the ventilation and the street set a floor you can never go below and never need to. This is liberating news. Once your rack is within 3 dB of the room’s noise floor, further work is inaudible, and you can stop. My own target was 33 dB(A) at two metres against a 30 dB(A) floor, and hitting it ended the project.
Measure in a fixed spot at a fixed time, note the number, change one thing, measure again. Changing three things and measuring once teaches you nothing about which one worked, and acoustic work is full of interventions that sound like they helped because you spent a Saturday on them.
Treating what’s left
Once the hardware is as quiet as it will get, the enclosure becomes worth it.
A closed rack cabinet with acoustic lining. Purpose-built quiet racks exist and cost a fortune; a standard closed cabinet lined with 25 mm acoustic foam gets you most of the way for a fifth of the price. Expect 6–10 dB(A). The physics is unforgiving here: every hole leaks. An enclosure with a 5% open area performs barely better than one with 50% open area, because sound takes the path of least resistance and a cable gland is a very low resistance. Baffle the cable entries with foam-lined labyrinths and expect to spend an evening chasing gaps.
And now the thermal problem. A sealed cabinet becomes an oven. My cabinet ran 14°C above ambient with the door shut, which pushed the fan curve into the region I had just spent a weekend escaping, and made the rack louder than it was in open air. The fix is a lined intake and exhaust path — two 140 mm fans at 400 rpm moving air through a foam-lined S-bend, which absorbs the treble on the way out while letting the air through. It works, and it is fiddly, and it is the part every guide skips.
Decouple everything from the structure. Drives and fans transmit vibration into the chassis, the chassis into the rack, the rack into the floor, and the floor into the flat below. Sorbothane feet under the rack, rubber grommets on drive cages, and soft-mounted fans cost about €30 and are the best value in this entire article. Structure-borne noise is what your neighbours complain about, and no amount of foam touches it.
Placement matters more than treatment. A rack in a corner couples to two walls and gets a bass boost of several dB for free. The same rack a metre out into the room, on a rug, behind a bookshelf, is meaningfully quieter with zero equipment. Move it before you buy anything.
The topology consequence
Noise reshapes your architecture, which is the part that catches people out. In a basement you would happily run one loud, powerful box and virtualise everything on it. In a flat, the quiet answer is several small machines that each idle passively, and that changes your design: you now have a cluster whether you wanted one or not, with the storage question, the shared-state question and the “which machine does DNS resolve on when the other one is off” question all arriving uninvited.
I have found this to be a net positive, with one caveat. The upside is that fanless mini PCs force honest resource limits, and a service that cannot fit on a 16 GB node gets scrutinised rather than accommodated. The downside is that the storage layer still wants a machine with real drives in it, and that machine is the loudest thing you own. My compromise is a single drive-bearing NAS that is asleep most of the day, woken on a schedule, with the always-on services living on silent, diskless nodes that mount nothing from it. It took two rebuilds to arrive there and it is the arrangement I would start from now.
Troubleshooting
“It’s quiet, but there’s a hum I can’t locate.” Likely a transformer or a PSU coil at mains frequency and its harmonics, or a fan at blade-pass frequency. Record it on your phone and look at the spectrum: a sharp peak at 50/100 Hz is electrical, a peak at some multiple of (rpm ÷ 60 × blade count) is a fan. Electrical hum is a component swap; fan tonality sometimes disappears with a 50 rpm change, because you have moved off a resonance in the chassis panel.
Noise got worse after adding foam. You have closed the airflow path and the fans have ramped. Check your temperatures before you check your ears — this is the single most common self-inflicted wound in the hobby.
A fan is audible only occasionally, and it’s maddening. Hunting: the fan curve is oscillating around a threshold, ramping up, cooling, ramping down, repeatedly. Add hysteresis by widening the gap between MINTEMP and MAXTEMP, or increase INTERVAL so the loop reacts more slowly. A fan at a constant 700 rpm is far less noticeable than one cycling between 500 and 900, because your ear tracks change rather than level.
Measured dB(A) doesn’t match how annoying it is. Because A-weighting was designed for occupational hearing damage rather than for a bedroom at midnight. A tonal 38 dB(A) whine is far worse than 44 dB(A) of broadband whoosh. Trust your ears over the meter, and aim for noise that is broadband and steady.
The neighbours complain but you can’t hear it. Structure-borne transmission. Your floor is a large, efficient loudspeaker driven by your rack’s feet. Decouple, and check whether the rack touches a wall — a rack bolted to a stud wall turns the entire wall into a diaphragm.
The verdict
A genuinely silent rack in a flat is achievable, and the honest route is subtractive. You get there by choosing hardware that never needed to be loud, and the treatment is a modest refinement at the end. The order that works: delete the 1U kit, fix the fan curves, decouple from the floor, move it away from the corner, and only then think about foam and cabinets.
The uncomfortable conclusion is that noise should be a purchasing criterion with the same weight as core count. The used enterprise machine at €120 that idles at 78 W and 52 dB(A) is worse value than a €300 mini PC, once you price the electricity from the real cost of self-hosting and the fact that you will eventually move the loud thing into a cupboard where it will cook. I know this because I own the cupboard.
Who is this for? Anyone whose rack shares a wall with a person. Who should skip it? Anyone with a garage, a basement, or a genuinely tolerant household — go and buy the loud thing, you have earned it, and I am mildly jealous.
Related Content
