The full story
This is the episode's narration, word for word. Headings jump to that point in the video.
The question 0:00
Last time, two gates wired in a loop held a single bit. But that bit is lost when the power goes off. Your phone is different. Let the battery go flat, and your photos are still there next week. Yet the app you had open has to start again from scratch. So what does remembering even mean, for a chip?
Two ways to remember 0:24
A memory needs a state that stays put after whatever set it has gone, and most memory chips do it in one of two ways. One is to hold it actively, with circuits that keep pushing each other into place. The other is to trap something, a little packet of electric charge, and come back for it later. Memory that forgets without power is called volatile. Memory that keeps it is non-volatile.
Six transistors 0:53
The holding kind is static RAM, or SRAM. At its heart are two of last episode's inverters, each one's output wired to the other's input. If the left one outputs a 1, the right one is forced to output 0, which keeps the left one at 1. The mirror image is stable too, so the pair stores one bit. Add two switches to connect it to the outside world for reading and writing, and that's six transistors per bit.
Fast, but big 1:25
Why use it? Because SRAM is the fastest kind of memory cell there is. The cell drives its own answer, and it's built from the same transistors as logic, so it sits right beside the processor. The catch is size. Six transistors for every bit makes it the most expensive memory per bit, which is why the fast caches take up a surprisingly large share of a processor chip.
Why it forgets 1:54
And it forgets because the loop is only held up by the power supply. Cut the power, and nothing holds either side. "Static" just means it doesn't need topping up. It doesn't mean it survives a power cut.
An idea on the couch 2:10
Back in the nineteen-sixties, the main memory in most computers was woven from tiny magnetic rings. In 1966, an IBM engineer called Robert Dennard came home from a talk about shrinking that magnetic memory. On his living-room couch, he worked out that one transistor and one capacitor could hold a bit instead of six transistors. Yes, the same Dennard whose group wrote episode five's rules for shrinking transistors.
A bucket of charge 2:41
That's dynamic RAM, or DRAM. The capacitor is a tiny bucket for charge. Full means one, empty means zero. The transistor is the door. To write, you open the door and fill the bucket, or drain it. Close the door, and the charge is trapped.
Reading wipes it 3:03
Reading it is delicate, because the bucket is tiny, only a few millionths of a billionth of a farad. Open the door, and its charge spreads into the much bigger wire it's connected to, nudging the voltage by only tens of thousandths of a volt. A circuit called a sense amplifier magnifies that nudge into a clean one or zero. But the charge has spilled out, so every read wipes the cell, and the amplifier has to write it straight back.
The leak 3:34
And the bucket leaks. Charge seeps through the capacitor's insulator, through the junctions, and through the transistor, which is never perfectly off. Dennard's patent said it up front: the charge leaks off, so the memory has to be regenerated. So the chip reads and rewrites every row, roughly every sixty-four milliseconds. That's called refresh, and it's why the memory is called dynamic. Heat makes it leak faster, so above eighty-five degrees Celsius, a typical DDR4 chip refreshes twice as often.
Volatile, but not instant 4:11
Sixty-four milliseconds is set by the leakiest cells. Most hold on much longer: one study worked out that sorting rows by how long they hold could cut refreshing by about three-quarters. In 2008, a Princeton-led team showed memory keeping its data for seconds after the power went off. Chilled with an upside-down can of compressed air, typically fewer than one per cent of the bits faded in ten minutes. They used it to steal encryption keys.
Why DRAM won 4:44
One transistor and one capacitor take up far less room than six transistors, so DRAM gives you many more bits for the money. In 1970, Intel's 1103 chip held a thousand bits at about a cent each, and it soon began replacing magnetic memory. Funnily enough, it used three transistors per bit, not Dennard's one. His cell took over in the mid-seventies, and it's in essentially every computer's main memory today.
Memory that never changes 5:16
Some memory never needs to change at all. Read-only memory, or ROM, is written at the factory. The last mask in making the chip decides which crossings in a grid of wires are connected, and that pattern is the data. It's dense and cheap, and the first Super Mario Brothers cartridge alone was made over forty million times. Back in 1970 came PROM, which you could program yourself by burning tiny fuses, but only once.
The bug that became a feature 5:48
At Intel, in late 1969, an engineer called Dov Frohman was chasing a fault. Electrons were getting stuck in the glass insulation of one of the company's memory chips, and changing how it behaved. He realised that was a way to store a bit. The idea of a floating gate had already been described in 1967, by Dawon Kahng and Simon Sze at Bell Labs. That's the same Kahng we met in episode five.
Trapped in glass 6:22
A floating gate is an extra gate, buried inside the oxide, sealed in glass on every side, with no wire to it at all. Electrons put on it stay for years. Why can't they leak away? Remember silicon's glass skin from the first episode. To get into it, an electron would have to climb a step of about three electron volts. At room temperature, heat gives an electron only around a hundredth of that. So they stay put.
Reading the trapped charge 6:55
And you read it without touching it. Electrons on the floating gate cancel part of the control gate's pull, so it takes a higher voltage to turn the transistor on. Apply a test voltage in between. If the transistor switches on, the floating gate is empty. If it stays off, it's charged.
Erased by light 7:18
In 1971, Intel launched Frohman's chip, the 1702. You programmed it with electricity. To wipe it, you shone ultraviolet light through a little quartz window on top for twenty minutes or so, which gave the trapped electrons enough energy to escape. In the film Frohman showed at its debut, Gordon Moore recalled, the bits dropped out one by one, until only the Intel logo was left. Leave an EPROM in direct sunlight for about a week and it would be wiped too, so datasheets told you to stick a label over the window.
Through the wall 7:59
So how do electrons get through glass in the first place? One way is to make them hot. Strong voltages race electrons along the channel, and a few near the end gain enough energy to leap the wall. Or you make the glass only a few nanometres thick and put around twenty volts on the gate above it. In a field that strong, electrons tunnel straight through the barrier, a quantum effect, without ever having the energy to climb it. Reverse the field and they tunnel back out, so by the end of the seventies, chips could be erased with electricity, no window needed.
A flash of erasing 8:38
But those early chips needed two transistors per bit. At Toshiba, from 1980, Fujio Masuoka cut that to one, by erasing a whole block of cells at once. By most accounts, a colleague said the erase looked like a camera flash, and the name stuck. He presented flash memory in 1984. Then came NAND flash, with cells strung in long chains, like the transistors in last episode's NAND gate. Each chain shares its wiring, so it packs much tighter, and it's what fills SSDs, memory cards and phones.
More than one bit per cell 9:20
Then engineers got greedy: instead of charged or empty, they put a measured amount of charge on the gate. Four levels store two bits. Eight levels store three. Some flash today reads sixteen levels, four bits in one cell. The price is that the levels crowd together, so a little lost charge can blur one into the next.
Why it wears out 9:45
And every trip through the glass does damage. Each time electrons are forced through, some atomic bonds in the oxide break, leaving traps where charge can leak. So flash has a lifetime. Cells storing one bit are typically rated for around a hundred thousand write-and-erase cycles, and those storing four, around a thousand. That's why your phone and SSD quietly spread writes around, so no part of the chip wears out early.
Why not make it all flash? 10:17
So why isn't your computer's working memory flash? Because flash is slow to write and wears out. The nearest cache answers in about half a billionth of a second, and main memory in about a hundred billionths. Reading a page of flash can take twenty-five millionths of a second, hundreds of times longer, and writing takes longer again. So computers use a ladder: a little fast SRAM by the processor, holding what it used most recently, then lots of DRAM, then a huge, cheap store of flash that survives a power cut.
Logic plus memory 10:55
So now we have switches that decide, and silicon that remembers. Put logic and memory together, and you can build a machine that follows instructions. But there's more than one way to build it. A chip that can do anything, or one that does just one thing, very fast? CPU, ASIC, FPGA and GPU. That's next time.
Sources
Every factual claim in the episode is tied to one of these. Spotted an error? Tell us.
- Halbleiter.org: "Sequential Logic: Flip-Flops, Latches, and Clock Trees" (E07 bridge) — halbleiter.org
- Chenming Hu, Modern Semiconductor Devices for Integrated Circuits (Pearson, 2010),… — chu.berkeley.edu
- Halbleiter.org: "The Flash Memory Cell" (floating gate, tunnel oxide, HCI/FN,… — halbleiter.org
- Halbleiter.org: "The SRAM Cell" — halbleiter.org
- Halbleiter.org: "The DRAM Cell" (few fF, leakage, ~64 ms refresh, sense amp,… — halbleiter.org
- Halbleiter.org: "Processor Architecture" (registers, L1 32–64 KB, L2/L3, DRAM,… — halbleiter.org
- Computer History Museum, Silicon Engine: "1966: Semiconductor RAMs Serve High-speed… — computerhistory.org
- IBM History: "DRAM" (Dennard, 1966, couch; patent 1968; Intel 1-kilobit 1970; ferrite… — ibm.com
- R. H. Dennard, US Patent 3,387,286 "Field-effect transistor memory" (filed 14 Jul… — patents.google.com
- Micron Technology: 16Gb DDR4 SDRAM datasheet (MT40A4G4/2G8/1G16): "64ms, 8192-cycle… — mm.digikey.com
- J. A. Halderman et al., "Lest We Remember: Cold Boot Attacks on Encryption Keys",… — usenix.org
- J. Liu, B. Jaiyen, R. Veras, O. Mutlu, "RAIDR: Retention-Aware Intelligent DRAM… — pdl.cmu.edu
- Computer History Museum, Silicon Engine: "1970: MOS Dynamic RAM Competes with Magnetic… — computerhistory.org
- Computer History Museum, Silicon Engine: "1965: Semiconductor Read-Only-Memory Chips… — computerhistory.org
- Computer History Museum, Silicon Engine: "1971: Reusable Programmable ROM Introduces… — computerhistory.org
- Intel Virtual Vault: "A Success…Out of Quality Control Issues" (EPROM, Frohman, 1101,… — intel.com
- Chenming Hu, Modern Semiconductor Devices for Integrated Circuits, ch. 5 §5.1 (SiO₂… — chu.berkeley.edu
- HyperPhysics (R. Nave): "Fermi-Dirac Distribution / Fermi Level" ("kT is only about… — hyperphysics.gsu.edu
- Swissbit AG, Application Note AN2107en "NAND flash endurance testing" (2022, CC BY 4.0) — swissbit.com
- Semiconductor History Museum of Japan: "1984: Flash memory released (Toshiba)" (2019) — shmj.or.jp
- STMicroelectronics M27C256B 256 Kbit UV EPROM datasheet, "Erasure operation" (2537 Å,… — html.alldatasheet.pl
- IEEE Spectrum: "Chip Hall of Fame: Toshiba NAND Flash Memory" — spectrum.ieee.org
- Wikipedia: "Fujio Masuoka" (NOR 1984, NAND IEDM 1987, Ariizumi naming, 2006 settlement) — en.wikipedia.org
- R. Sheldon & C. Sliwa, "What is NAND flash wear-out?", TechTarget, 14 Mar 2022 (P/E… — techtarget.com
- P. Norvig, "Teach Yourself Programming in Ten Years": "Approximate timing for various… — norvig.com
- KIOXIA TC58NVG1S3HTA00 2 Gbit SLC NAND datasheet (tR 25 µs max; tPROG 300 µs typ.;… — wmsc.lcsc.com
Researched and scripted with AI assistance, fact-checked claim by claim, with synthetic narration and diagrams drawn in code. How we make episodes.
