Hard Disk Drive: A Head Flying Nanometers Above a Spinning Platter

Inside the Sealed Box: Platters, Spindle, Actuator and Heads

A 3.5-inch drive is a sealed aluminium casting 101.6 mm wide and 26.1 mm tall. Inside is a stack of platters 95 mm across, made of aluminium alloy or glass. Each face carries sputtered thin films: a soft magnetic underlayer, a ruthenium interlayer and a recording layer of cobalt-chromium-platinum grains under ~10 nm across, kept apart by oxide. On top sit a carbon overcoat a couple of nanometres thick and ~1 nm of perfluoropolyether lubricant.

A brushless DC spindle motor turns the stack on a fluid-dynamic bearing, whose grooves pump an oil film into a load-carrying wedge. Fluid bearings replaced ball bearings in the early 2000s because they run quieter with far less unpredictable wobble (non-repeatable runout).

A rotary actuator swings a comb of arms on a ball-bearing pivot, carrying one head for every recording surface. Each head is a slider on a stainless-steel suspension and gimbal. This is a femto-format block of alumina–titanium carbide, 0.85 × 0.70 × 0.23 mm, with the writer, reader and a tiny heater built as thin films on its trailing edge. A flat voice coil behind the pivot drives the arm. Air-filled drives breathe through a filtered hole; helium drives are hermetically sealed.

Step by Step: How the Drive Reaches a Sector

1. Spin up. The spindle reaches 7,200 rpm = 120 revolutions per second (one turn every 8.33 ms). At the outer tracks, ~46 mm from the axis, surface speed is v = 2πrf = 2π × 0.046 m × 120 s⁻¹ ≈ 35 m/s (~125 km/h).

2. Load. In most modern drives, the heads slide off a plastic ramp onto the surface once it is at speed.

3. Fly. Rails etched into the slider's air-bearing surface compress gas dragged along by the disk into a film whose lift balances a ~2 gf suspension preload. The slider pitches nose-up, so its trailing edge, where the read and write elements sit, is the lowest point, of order 10 nm up.

4. Close the gap. Before reading or writing, a thermal fly-height-control heater expands the surrounding thin films, bulging the read/write element down to ~1–3 nm clearance. The bulge is withdrawn between operations to cut wear.

5. Find the track. Hundreds of factory-written servo wedges per revolution each hold a sync mark, a Gray-coded track number and burst patterns that give a position error signal: the head's offset from track centre.

6. Seek and settle. Current in the voice coil accelerates the arm, then reverses to brake it. A track-following loop then holds the head over tracks ~50–60 nm apart.

7. Wait. The sector is on average half a turn away: 4.17 ms.

8. Write or read. To write, a coil drives flux from a narrow main pole straight down through the recording layer and back through the soft underlayer, setting grains up or down (perpendicular recording). To read, a tunnel-magnetoresistance sensor changes resistance as the field from passing magnetisation transitions tilts its free layer. The read channel decodes the noisy signal with LDPC error correction.

9. Park. At idle or power-down the heads retract onto the ramp. After a sudden power loss, the still-spinning spindle motor's rectified back-EMF powers the retract.

The Air Bearing: Why the Head Flies, Worked Through

The slider is a self-acting gas bearing, the gas cousin of an oil journal bearing. Viscosity drags gas with the moving disk into the narrowing gap under the slider, raising its pressure. For a continuous gas, the pressure p obeys the compressible Reynolds equation:

∂/∂x(p h³ ∂p/∂x) + ∂/∂y(p h³ ∂p/∂y) = 6µU ∂(ph)/∂x

Here h is the local gap, U the disk speed and µ the viscosity. But air molecules travel ~65 nm between collisions (the mean free path λ). Under a slider ~10 nm high, the Knudsen number Kn = λ/h is about 6.5; at a 2 nm clearance it is about 33. The molecules slip along the walls, so pressure-driven flow leaks out of the film more easily and it builds less pressure than continuum theory predicts. The first correction was a slip-flow factor, h³(1 + 6Kn) (Burgdorfer, 1959). Modern simulators use the Fukui–Kaneko (1988) generalised lubrication equation, whose flow factors come from the linearised Boltzmann equation and hold at any Knudsen number.

Worked example: how hard does the film push? A 2 gf preload is 0.002 kg × 9.81 m/s² ≈ 19.6 mN. Over the 0.85 × 0.70 mm footprint (0.595 mm²), that averages 19.6 × 10⁻³ N ÷ 0.595 × 10⁻⁶ m² ≈ 33 kPa above ambient, a third of an atmosphere. Real designs add a recessed pocket of sub-ambient pressure that sucks the slider down, so the small trailing pad carries far more than the average. The push–pull makes the bearing stiff: pressure climbs sharply if the gap shrinks and falls if it widens, so the slider follows the disk's waviness instead of hitting it. A 1 µm dust particle, by contrast, is ~500 times taller than a 2 nm gap.

Rarefaction sets practical limits too. At altitude the mean free path lengthens and the head flies lower, so air-filled drives are typically rated to operate only up to ~3,000 m. Helium has roughly the viscosity of air but 1/7 its density, so it still makes lift with far less turbulence. Its ~175 nm mean free path demands its own bearing designs.

Seek, Latency and Throughput: The Access-Time Arithmetic

The actuator is a voice-coil motor. A flat coil sits between neodymium-iron-boron magnets. Its two active legs carry opposite currents over opposite magnetic poles, so both push the same way. Each leg feels F = NBIL (turns × flux density × current × length), so torque is proportional to current and reverses with it.

Seek. On a long move the servo accelerates hard for half the distance and brakes for the rest. At constant acceleration a, a distance d takes t = 2√(d/a). Two random tracks are on average a third of the stroke apart. Averaging √d over random pairs gives 8/15 of the full-stroke time, so in this idealised model an average seek takes about half as long as a full-stroke seek. For a 7,200 rpm drive the average is ~8.5 ms.

Track following. Tracks ~50–60 nm apart mean roughly 420,000–510,000 tracks per inch. The head must typically stay within about a tenth of the pitch, ~5 nm, despite runout and vibration. Many high-capacity drives add a piezoelectric microactuator as a fast second stage.

Latency. Half of one 8.33 ms turn (60 s ÷ 7,200) is 4.17 ms. A 15,000 rpm enterprise drive waits 2.00 ms.

Worked example: random I/O. Seek plus latency is 8.5 + 4.17 ≈ 12.7 ms, so a 7,200 rpm drive serving one request at a time manages about 1 ÷ 0.0127 s ≈ 79 random reads per second (command queueing that reorders requests by rotational position can do somewhat better). Sequential reads are far quicker: with bits packed at roughly constant linear density (zoned recording), outer tracks hold about twice the data of inner ones, and a modern high-capacity drive streams ~250 MB/s or more there.

From RAMAC to HAMR: Seven Decades of Hardware

• IBM 350 RAMAC (1956): 50 × 24-inch platters, 3.75 MB, 1,200 rpm, 2,000 bits/in². Compressed air held the heads off the disk.
• IBM 1301 (1961): the first heads to fly on a self-acting air bearing.
• IBM 3340 “Winchester” (1973): platters and low-load heads sealed together, with the heads landing on the lubricated disk when it stopped. Its planned 30 MB + 30 MB configuration echoed the Winchester .30-30 rifle.
• Read heads: magnetoresistive (IBM, 1991), then giant magnetoresistance (IBM Deskstar 16GP, 1997; Nobel Prize for Fert and Grünberg, 2007), then tunnel magnetoresistance from ~2005.
• Perpendicular recording: first shipped by Toshiba in 2005.
• HGST Ultrastar He6 (November 2013): helium allowed thinner platters and 6 TB on seven platters in the standard 26.1 mm-tall case.
• Seagate Mozaic 3+ (January 2024): heat-assisted magnetic recording (HAMR), 3 TB or more per platter.

Why HAMR was needed. A grain keeps its magnetisation only if its anisotropy energy (anisotropy constant × grain volume) is at least ~40–60 times the thermal energy kT. Below that, heat flips grains at random and data fades: the superparamagnetic limit. Smaller grains need a higher anisotropy constant, but the head's writing field is limited. HAMR uses iron-platinum (FePt), whose anisotropy is roughly an order of magnitude higher than that of the cobalt alloys. A laser diode on the slider feeds light to a gold near-field transducer, which heats a spot tens of nanometres across to ~450 °C. That is near FePt's Curie point, so the writer can set the grains, which cool and lock in nanoseconds. Areal density has risen from RAMAC's 2,000 bits/in² to over 1 terabit per square inch, more than 500 million-fold.

Specifications, Testing and How Drives Fail

Specifications. Consumer drives quote an unrecoverable read error rate of 1 in 10¹⁴ bits (one error per ~12.5 TB read); enterprise drives quote 1 in 10¹⁵. Enterprise drives also carry workload ratings around 550 TB/year and an MTBF that is commonly 2.5 million hours. MTBF is a fleet statistic, not a lifespan: 8,760 h ÷ 2,500,000 h is a 0.35% annualised failure rate. Backblaze's data from its fleet of hundreds of thousands of drives typically shows ~1–2% failing per year. Standards include the SFF-8301 3.5-inch form factor, SATA and SAS interfaces, and 4,096-byte “Advanced Format” sectors, adopted around 2011.

Testing. Platters are glide-tested with a head carrying a piezoelectric sensor that detects bumps. Each head's heater is calibrated by raising its power until a contact sensor or the read-back signal detects touchdown, then backing off. In service, SMART (Self-Monitoring, Analysis and Reporting Technology) logs reallocated sectors and errors. But a 2007 Google study of more than 100,000 drives found that over a third of failed drives had shown no SMART warning.

• Head crash: a particle, lubricant pick-up or a shock during operation brings the slider into contact. It gouges the overcoat and magnetic layer, and the debris spreads the damage. Operating shock ratings are far lower than non-operating ones, because a parked head is safe on its ramp.
• Vibration: in 2008 engineer Brendan Gregg filmed disk latency spiking as he shouted at a drive array. In 2016 a fire-suppression gas discharge in ING Bank's Bucharest data centre was loud enough to damage drives and knock services offline.
• Firmware: in 2009 a bug left some Seagate Barracuda 7200.11 drives unresponsive at power-up, with the data intact but unreachable.
• Environment and misuse: above the rated altitude, fly height drops. Many helium drives report a helium-level SMART attribute to flag leaks. In 2020 some WD Red NAS drives turned out to use drive-managed shingled magnetic recording (SMR), and some RAID and ZFS rebuilds slowed dramatically or failed.

Not an SSD, Not a Record Stylus

Versus an SSD. A solid-state drive has no moving parts. It stores charge in NAND flash cells and reaches any block electronically in well under a millisecond, so it wins at random I/O by orders of magnitude and survives shocks. But each cell survives only a limited number of program/erase cycles, and unpowered cells slowly leak charge. A hard drive's grains have no write-cycle limit and hold data for years without power, and hard drives cost several times less per terabyte.

Versus a record stylus. A turntable stylus is meant to touch. It rides in the groove at ~1–2 g tracking force, moving at about 0.5 m/s on the outer groove of a 33⅓ rpm LP. A disk head carries a similar ~2 gf preload but rests entirely on gas at ~35 m/s on a 7,200 rpm drive's outer tracks, and any touch is a head crash. Floppy disk heads did run in contact, which is why floppies wore out.

Hard disk drive versus the storage devices it is confused with

Device	What stores the data	Does anything touch the medium?	Typical random access
Hard disk drive (3.5-inch, 7,200 rpm)	Magnetised grains in a cobalt-platinum-alloy film on glass or aluminium platters	No: the head flies ~1–3 nm above the surface; contact is a head crash	~12.7 ms (~8.5 ms seek + 4.17 ms latency)
SSD (NAND flash)	Electric charge trapped in flash memory cells	Nothing moves	Well under 1 ms
3.5-inch floppy disk	Magnetic particle coating on a flexible plastic disk	Yes: the heads press on the medium at 300 rpm	~0.2 s
Vinyl record player	Analogue wiggles cut into a spiral groove	Yes: the stylus rides in the groove at ~1–2 g tracking force	Not random access: the needle is placed by hand
LTO tape cartridge	Magnetised tracks on ~1 km of tape	Yes: the tape runs across the heads	Tens of seconds to wind to a file
IBM 350 RAMAC (1956)	50 × 24-inch platters, 3.75 MB, 1,200 rpm, 2,000 bits/in²	No: heads held off the disk by compressed air	Hundreds of milliseconds