Why Didn't Hard Drives Change in 80 Years?

Hear that rhythmic clicking sound?

It's the sound of death.

Specifically, your data dying.

If you used computers in the late 90s, I apologize. I should have given you a trigger warning. And if you have no idea what I'm talking about, there was a time when that sound meant your entire semester's homework, or all the records for your small business, were about to go up in smoke.

The Iomega Zip drive looked like a GameCube and a floppy disk reader had a baby. Back in the day, your school projects, design files, and work documents could all live on one. It held 100 megabytes at launch, which sounds pitiful now, but in 1995 that was a lot. One Zip disk held about as much as 69 standard floppy disks.

But if your Zip drive started clicking, things were about to go terribly wrong.

The "click of death" was so infamous that some Iomega customers sued the company in 1998, claiming the drives damaged their cartridges and made the data impossible to recover. Iomega disputed how widespread the problem was, but the sound stuck in people's heads because it was such a visceral representation of how fragile data storage could be.

Specifically, the fragility of magnetic storage.

For decades, computers improved in every direction except one. CPUs measured work in nanoseconds. RAM answered queries in tens of nanoseconds. Networks went from slow telephone-based modems to cable broadband. But hard-drive storage still had to wait for a tiny mechanical arm to move across a spinning plate.

In the 2000s, the hard drive was the component that made your computer feel slow more than anything else. It's why a relatively cheap SSD upgrade could make a seven-year-old gaming rig feel new again. And if you're a developer, it's why some database optimizations you absolutely had to know about in 2005 are no big deal today.

So, how did the hard drives of old work? How can their solid-state replacements store bits without any physical movement? And what do the massive hardware improvements of the last 20 years mean for how we should write fast software?

The First Hard Drive

In 1956, IBM introduced the RAMAC 305, the first commercial computer system built around a random-access disk drive. Its storage unit, the IBM 350, was five feet tall, six feet wide, weighed over a ton, and had fifty 24-inch platters spinning at 1,200 revolutions per minute.

The thing was massive. And the storage capacity?

Almost 4 megabytes, if you convert those old six-bit characters into modern eight-bit bytes.

For comparison, one uncompressed 4K frame is about 25 MB: roughly six times more data. Storing a whole modern video on RAMACs would require a lot of real estate.

But however ridiculous it seems today, RAMAC solved a real problem in the 50s.

Before disks, data usually lived on punch cards, magnetic tape, or paper. Magnetic tape is great if you want to read stuff in order. It's not great if you want to update user 825's full name right now. You have to wind through the tape until you reach that exact record. It's like fast-forwarding a VHS to find one scene from The Princess Bride.

RAMAC's big breakthrough was random access. Instead of walking through every record in sequence, the machine could move a read/write head straight to the part of the disk where the data lived, like picking up a record player's needle and moving it an inch. For commercial disk storage, that was brand new in 1956. It let businesses pull up one account, one invoice, or one inventory record directly.

But the breakthrough created the bottleneck that haunted computing for the next 50 years.

How a Hard Drive Works

A hard drive has a physical plate with data encoded on its surface, spinning thousands of times per minute.

The disk contains tiny magnetic regions. Depending on each region's magnetic direction, the drive interprets it as a one or a zero. A read/write head sits above the plate on a movable actuator arm. The head can detect and change those magnetic fields without touching the plate.

The "click of death" is the drive trying and failing to position its head over the right spot. The click isn't the original problem. It's the actuator snapping through a recalibration motion after the drive fails to complete an operation, usually because the disk media or mechanism is damaged.

Modern hard-drive heads ride on a cushion of air created by the spinning disk, floating unbelievably close to the platter. In the original IBM 350, the heads rode about 800 microinches above the surface. A microinch is one millionth of an inch, so the head floated 0.0008 inches above the plate without touching it.

Today, we use nanometers to measure the gap. The IBM 350's heads flew about 20,000 nanometers above the disk. Modern drives have shrunk that distance to just a few nanometers.

Brilliant engineering, especially for the 50s. But it had a big problem: latency, the time between asking a device for data and getting it back.

To read data, the drive does two physical things. First, the arm seeks the right track. Then the platter rotates until the correct sector passes under the head. Computer architecture classes still teach disk access around these mechanical costs because that delay was often the worst bottleneck when loading a new game or opening a large program.

The IBM 350 could take hundreds of milliseconds, or nearly a full second, to read or write data. Even much later hard drives commonly lived in the millisecond world for random access. Harvard's CS 61 storage notes give 12 milliseconds for an external hard drive and about 4 milliseconds for a high-end 2020-era disk.

Ten milliseconds feels instant to a human. To a CPU, it's an eternity. At 4 gigahertz, that's about 40 million clock cycles. Your CPU could perform millions of simple operations while it waits for the drive to find one chunk of data.

That's why old computers felt slow in specific ways.

Open Photoshop? Go get a coffee.

Launch a game? Better take a quick bathroom break.

You weren't waiting for the CPU to think. You were waiting for data from disk.

Hard-drive companies improved the engineering. Platters got smaller. Tracks got denser. Caches got smarter. But the fundamental problem remained: a motor spun a disk while an arm hunted for the right spot.

Two Times Faster vs. 9,000 Times Faster

It wasn't until the Windows Vista and Windows 7 era that things really changed. You could take the same machine, with the same CPU, RAM, and operating system, swap only the storage drive, and make it feel like a whole new computer.

Over the last 20 years, high-end CPU clock speeds have gone from about 3 GHz to 6 GHz. There have been plenty of other improvements, but raw clock speed has only about doubled.

Storage is a different story. In 2004, StorageReview tested Western Digital's 10,000 RPM Raptor and measured one random read taking 7.8 milliseconds. Two decades later, a 2 TB WD_BLACK SN850X solid-state drive is rated for 1.2 million random reads per second.

That's not twice as fast. That's a jump from about 128 reads per second to 1.2 million: an improvement of more than 9,000 times.

The primary reason is the solid-state drive. No platter. No flying head. No clicking.

Just chips.

But if a hard drive stores bits by changing the magnetic direction of tiny regions on a plate, how does an electronic chip store a file even when the power goes off?

Flash Memory and NAND

In the 1980s, Toshiba engineer Fujio Masuoka helped invent flash memory. He presented his flash-memory design in 1984, which prompted Intel to develop NOR flash. In 1987, Masuoka and his colleagues presented NAND flash at the IEEE International Electron Devices Meeting.

That word, NAND, appears on almost every SSD spec sheet: NAND flash, 3D NAND, TLC NAND.

With flash memory, data lives in tiny transistor cells. Each cell has an insulated pocket where electrons can get stuck. If electrons are trapped there, the transistor won't conduct current and represents a zero. If they aren't, it conducts current and represents a one. Those electrons stay trapped even when the power turns off. That makes flash non-volatile memory, unlike RAM, which forgets everything when the power dies.

The NAND part stands for "NOT AND," one of the basic logic gates computers are built from. A NAND gate takes two binary inputs and produces one output. It's an AND operation followed by a NOT operation.

Input A ----\
             AND ---- NOT ---- Output
Input B ----/

An AND operation outputs one only when both inputs are one. NAND flips that result.

    | AND | NAND
0 0 |  0  |  1
0 1 |  0  |  1
1 0 |  0  |  1
1 1 |  1  |  0

NAND flash got its name because its memory cells connect in long series strings that resemble the transistors in a NAND logic gate. A simple five-cell NAND string looks like this:

bit line
  |
[cell]-[cell]-[cell]-[cell]-[cell]
  |
source line

A comparable NOR architecture looks more like this:

bit line
   |
+-------+-------+-------+-------+
|       |       |       |       |
[cell]  [cell]  [cell]  [cell]  [cell]
|       |       |       |       |
source  source  source  source  source

NAND is denser per cell, making it better for storing as much data as possible. It takes less silicon for the same number of bytes. NOR is better at reading individual memory locations quickly, but it's slower to write and erase.

In a simple single-level-cell NAND string, the chip reads one cell by turning the others "on" with a higher pass voltage so they conduct regardless of what they're storing. It then checks whether the selected cell lets current flow.

No arm moves and no platter rotates. That's what "solid state" means.

The drive asks a grid of transistor cells: which of you are holding electrons?

The Tiny Computer Inside an SSD

Flash had one small problem that kept it from immediately killing the hard drive.

Writing damages it.

Every program-and-erase cycle stresses the insulating layer that traps charge. Hard drives can rewrite the same physical area again and again. Use flash cells enough, and they become unreliable.

So how do you build a hard-drive replacement from cells that slowly die whenever you use them?

The SSD presents one simple version of itself to your operating system, then does something much messier underneath.

From the operating system's point of view, an SSD looks like a normal block device. The OS says, "Write this kilobyte to block 123," and later asks, "Give me back the kilobyte from block 123." The SSD returns the data. Easy.

But not on the inside.

The controller keeps a secret map called a flash translation layer. To the operating system, block 123 is one stable address. But the controller may physically write that data somewhere else, then update its private map so future reads still work. This lets flash pretend to behave like a normal hard drive.

Flash writes in small pages, often 4 KB to 16 KB, but erases in larger blocks made of many pages. If one tiny page changes, the SSD may need to copy the still-valid pages elsewhere before it can free the block.

This cleanup is called garbage collection, not to be confused with the garbage collection performed by the Python or JavaScript runtime.

Because individual cells wear out, the SSD also uses its translation layer to spread writes across the drive so all the cells age more evenly. This process is called wear leveling.

The SSD is physically simpler than a hard drive because it has no moving parts. Logically, it's much more involved. It has a tiny computer inside while pretending to be a simple disk.

Cheaper, Denser, and Faster

That tiny computer used to be expensive.

In 2008, Ars Technica reported an MSRP of $1,099 for a 64 GB OCZ SSD, or about $16 per GB. Our World in Data puts inflation-adjusted hard-drive storage in 2008 at about 12 cents per GB. Spinning disks cost about one one-hundredth as much.

SSDs are absurdly cheaper now, but hard drives still win on raw capacity per dollar. In a May 2026 spot check, the cheapest new internal 3.5-inch hard drive on DiskPrices cost a little over 4 cents per GB, while the cheapest new internal SSD cost a little over 31 cents per GB. The hard drive remained about eight times cheaper.

For years, manufacturers made flash cheaper by packing more bits into each cell. But when each cell holds more states, the voltage gaps between those states narrow, making performance and lifespan harder to preserve.

Engineers needed another direction to grow.

Imagine planning a city on an island. You can't keep expanding outward, so you build up. 3D NAND did the same thing by stacking memory cells vertically. In 2013, Samsung started mass-producing 24-layer vertical NAND to break through the limits of flat, planar designs.

That helped make large SSDs cheap enough for normal people. Then the connection to the computer became the bottleneck.

Most early consumer SSDs connected through SATA, or Serial Advanced Technology Attachment. SATA was designed for disk drives, which made sense when disk drives were all anyone used. But when SSDs started delivering data at far lower latency, that old interface became a problem.

Its replacement was NVMe, Non-Volatile Memory Express. NVMe was designed for SSDs connected over PCI Express, the same high-speed connection family used by GPUs.

AHCI, the host-controller interface commonly used with SATA, was built around one command queue holding up to 32 commands. NVMe supports tens of thousands of queues, each holding tens of thousands of commands.

A command is a request such as "read 4 KB from address 1056." Queues are waiting lines for those commands. NVMe's design lets multicore CPUs keep a huge number of requests in flight and collect responses as quickly as an SSD can produce them.

Once storage became hundreds or thousands of times faster, old software assumptions started breaking down.

SSDs Changed How We Write Software

For decades, database design had to respect the hard drive. Sequential reads were fast. Random reads were slow.

If a database read every row in a table from front to back, the hard drive could stream the data once its head was in position. If an index sent the drive jumping from record to record, the actuator arm had to seek again and again.

PostgreSQL is a popular open-source database that's been around since the 90s, and you can still see that limitation in its query-planning configuration.

A database index is a data structure that can make looking up a row by its ID tens of thousands of times faster. But on spinning disks, an index can sometimes be a trap. The index tells the database exactly which rows to read, but if those rows are spread across the disk, the drive spends its time hunting with the actuator arm. Reading a larger chunk of the table in order may be faster.

SSDs changed that math.

Random-read-heavy workloads became more attractive. Caches, logs, compaction strategies, and storage engines all gained new tradeoffs, mostly better ones. Software architecture is still full of assumptions inherited from hardware that many production machines no longer use.

The same change is obvious on a personal computer. Replacing a hard drive with even a cheap SATA SSD is one of the lowest-cost ways to make an old gaming PC feel less miserable. It usually won't raise your average frame rate because that's mostly the GPU, CPU, and RAM. But it will cut loading and startup times.

The Catch: SSDs Forget

People hear "no moving parts" and translate it to "nothing can break."

No.

Flash stores electrical charge. And charge leaks.

A powered-off SSD doesn't hold data forever. Retention depends on temperature, workload, and wear. Under the JEDEC industry standard, a client SSD only has to retain data for one year at 30 degrees Celsius after reaching the end of its rated life, according to Western Digital's endurance white paper.

That doesn't mean every SSD forgets your files after one year. But flash retention isn't infinite, especially after wear.

SSDs trade mechanical decay for electrical decay.

With a hard drive, data can die because a head touches the platter. With an SSD, data can die from natural causes. The answer is simple: make backups.

And guess which kind of storage is great for rarely used, massive-scale data?

Spinning disks.

Backblaze buys drives by the truckload for cloud storage. It tracked hard-drive costs falling from over 11 cents per GB in 2009 to about 1.4 cents in 2022. HDDs remain the economic workhorse for bulk storage.

Cold backups, archives, surveillance footage, and giant media libraries still use physical disks. But for latency-sensitive computing, whether on a cloud server or your local gaming rig, SSDs have mostly won.

The hard drive didn't disappear. It settled into the jobs where being cheap matters more than being fast. The SSD took everything else, then quietly rewrote the performance assumptions underneath modern software.

Bibliography

1. Tony Smith, "Iomega sued over Zip-zapping 'click of death'." The Register, November 11, 1998.
2. IBM, "RAMAC." IBM History.
3. Computer History Museum, "1956: First commercial hard disk drive shipped."
4. Harvard CS 61, "Storage."
5. Fujio Masuoka, Masaki Momodomi, Yoshihisa Iwata, and Riichiro Shirota, "New ultra high density EPROM and flash EEPROM with NAND structure cell." IEEE IEDM, 1987.
6. KIOXIA, "What is NAND Flash Memory?"
7. IBM, "What is Flash Memory?"
8. KIOXIA, "Technology Development History."
9. Nitin Agrawal, Vijayan Prabhakaran, Ted Wobber, John Davis, Mark Manasse, and Rina Panigrahy, "Design Tradeoffs for SSD Performance." Microsoft Research / USENIX, 2008.
10. Jon Stokes, "SSDs in 2008: fast speeds (200MB/sec) over price cuts." Ars Technica, March 2008.
11. Our World in Data, "Historical price of computer memory and storage."
12. Samsung, "Samsung Starts Mass Producing Industry's First 3D Vertical NAND Flash." Samsung Global Newsroom, August 2013.
13. NVM Express, "About."
14. Cisco, "What is NVMe?"
15. PostgreSQL Documentation, "Query Planning."
16. Sean Webster, "SSD vs HDD Tested: What's the Difference and Which Is Better?" Tom's Hardware, 2019.
17. Western Digital, "SSD Endurance and HDD Workloads."
18. Andy Klein, "Hard Drive Cost Per Gigabyte." Backblaze, November 29, 2022.
19. Yuefan Xie and David B. Bogy, "Study of Head Disk Interface Characterization using Touchdown Sensor, Acoustic Emission Sensor, and Thermal Fly-height Control." Computer Mechanics Laboratory, University of California, Berkeley, 2015.
20. DiskPrices, "Disk Prices (US)."
21. PCMag Encyclopedia, "Zip disk."
22. KIOXIA, "What is Multi-level Cell Technology Realizing Larger Capacity Flash Memory?"
23. PCMag Encyclopedia, "Floppy disk."
24. Colin Scott, "Latency Numbers Every Programmer Should Know By Year." University of California, Berkeley.
25. PCMag Encyclopedia, "V.90."
26. PCMag Encyclopedia, "Cable modem."
27. Intel, "Intel Core i9 processor 14900."
28. Intel, "Intel Optane SSD 900P Series."
29. StorageReview, "Western Digital Raptor WD740GD." January 28, 2004.
30. SanDisk, "WD_BLACK SN850X NVMe SSD."
31. Amazon, "Western Digital 6TB WD Purple Surveillance Internal Hard Drive HDD."
32. Amazon, "Samsung SSD 9100 PRO with Heatsink 8TB."
33. Brian Santo and IEEE Spectrum staff, "Chip Hall of Fame: Toshiba NAND Flash Memory." IEEE Spectrum, June 30, 2017; updated 2025.
34. Codrut Neagu, "SSD vs. HDD: Do SSD drives give you higher frame rates in games?" Digital Citizen.
35. Zijian Zhao et al., "Paving the Way for Pass Disturb Free Vertical NAND Storage via A Dedicated and String-Compatible Pass Gate." arXiv:2403.04981, 2024.