Making RAM at Home — shantum's notes

ELI5/TLDR

A guy turned a shed in his backyard into a proper clean room and made his own computer memory from scratch. He built tiny switches paired with tiny batteries on a slice of silicon, charged them up, and read the charge back. It works. Not well enough to run anything useful yet, but the thing stores a bit of data, which is the whole job of RAM.

The Full Story

Why this is nuts to begin with

RAM prices are painful right now, and the reason comes down to three companies: Micron, Samsung, and SK Hynix. That’s it. New factories cost billions and take years. So Dr.Semiconductor did the most unreasonable thing possible — he built the tools in a backyard shed rated as a “class 100” clean room (fewer than 100 dust particles per cubic foot, which sounds lax until you remember your kitchen has millions).

What RAM actually is, when you zoom in far enough

Picture a spreadsheet with tens of thousands of rows and columns. At every intersection sits two things: a tiny switch (the transistor) and a tiny rechargeable battery (the capacitor). Flip the switch on, current flows in, battery charges up. That stored charge is a “1”. No charge is a “0”. Flip the switch off and the charge is trapped.

To read the bit, you flip the switch on again and let the charge flow back out. A detector at the other end notices it. Problem: reading drains the battery. So the chip has to constantly go around topping everything up, thousands of times a second. That’s the “D” in DRAM — Dynamic, as in “please keep refreshing me or I forget everything.”

Building it, one floor at a time

Chips are built like a layer cake. You start with a raw silicon wafer and cleave it into chip-sized pieces along its natural crystal planes — silicon breaks at clean 90-degree angles, almost obligingly.

Then you rust it. Literally. The silicon goes into a furnace at 1,100°C (volcanic-lava hot) where oxygen creeps into the surface and turns the top layer into glass. This glass is both protection and a stencil mask for later steps.

Photography, but for atoms

The patterning process is a kind of photography. You coat the chip with a light-sensitive goo (photoresist), shine UV through a mask with your design cut out of it, and the exposed goo becomes chemically vulnerable. Dunk the chip in developer solution and the exposed parts dissolve away. What’s left is your design, etched in goo, acting as a mini shield for what happens next.

“UV light generates a photo acid which when you take the sample and put it in a developer solution which is basic. The acid and base neutralize each other generating a salt which is dissolved away.”

A microscope-stepper system shrinks the mask pattern down so features end up smaller than a micron — below the size of a red blood cell.

Making the switch parts conductive

A transistor is a switch with three terminals: two you feed current through (source and drain) and one that controls whether the current flows (gate). To make the source and drain conductive, you have to “dope” the silicon — shove phosphorus atoms into it.

The industry does this with ion implantation machines, which cost as much as a mansion. Dr.Semi uses a cheaper trick: “phosphorus-doped spin-on glass.” You paint a phosphorus-laced liquid onto the chip, bake it, then heat the whole thing so the phosphorus diffuses from the glass down into the silicon underneath. The heat profile matters — the shallower and more controlled the diffusion, the smaller your features can be.

The gate — the actual switchy bit

Between the source and drain sits the channel. On top of the channel sits a very thin layer of glass — the gate oxide — and on top of that, the metal that receives the control voltage. When you apply voltage to the gate, it creates an electric field that either allows or blocks current from flowing through the channel below. That’s the switch.

Thinner gate oxide means better control, so this layer gets a special lower-temperature growth step at 950°C for 38 minutes, producing 20 nanometers of glass. Before that step, there’s a “piranha clean” — an acid so aggressive it dissolves skin. Named, cheerfully, after the fish.

Painting with atoms

The final step is depositing metal for the wires and contacts. Think of the chip with its goo pattern as a spray-paint stencil. Except instead of paint, you fire argon atoms at a slab of aluminum inside a vacuum chamber, which knocks aluminum atoms loose. Those atoms drift down and coat everything — the chip, the stencil, all of it. Then you dissolve away the goo, and the aluminum sitting on top of it peels off with it, leaving only the wires where you wanted them. This is called “lift-off.”

Does it work?

Yes, mostly. Tested on a semiconductor parameter analyzer with fine-needle probes, the transistor switches. The capacitor stores 12.3 picofarads, close to the theoretical ideal.

But there are compromises. The transistor has a flaw called “punch-through” — because source and drain are less than a micron apart, high voltages cause them to electrically merge and the gate loses control. Fine at low voltages, broken at high ones. This is the exact scaling problem real fabs spend decades fighting.

The capacitor leaks. Commercial DRAM holds charge for 64 milliseconds between refreshes. His holds it for just over 2 milliseconds. So his RAM would need to refresh roughly 30 times more often. It works — it’s just thirsty.

Key Takeaways

DRAM is literally a grid of tiny switches and tiny batteries; that’s the entire architecture.
Chips are built one layer at a time, like a sandwich — grow oxide, pattern it, etch, repeat.
Photoresist + UV + masks is photography at the nano scale; the same logic as a film negative.
Doping = injecting specific atoms into silicon to change its conductivity. Spin-on glass is the poor-man’s alternative to an ion implanter.
A transistor’s “gate oxide” is a thin glass insulator; thinner is better for control but harder to make reliably.
“Punch-through” is the short-channel effect where source and drain electrically merge when they’re too close — one of the core reasons shrinking transistors is hard.
Commercial DRAM refreshes every 64 ms; his shed-built version lasts 2 ms. The ratio tells you how much engineering separates an amateur result from a production one.
The entire global DRAM supply depends on three companies. That’s the economic fact that started this whole project.

Claude’s Take

Score: 9/10. This is the most satisfying kind of hardware video — the kind where someone does the impossible thing and also explains every step clearly enough that you understand why it’s impossible. The pacing is calm, the analogies (stencil for lift-off, rust for oxide growth, dimmer switch for transistor behavior) are genuinely useful, and the numbers he reports are the right numbers to report. He doesn’t oversell the result. The capacitor leaks 30x faster than commercial DRAM. He says so. The transistor has punch-through. He says so.

The only reason this isn’t a 10 is that it’s a progress report, not a conclusion. He’s built a few cells, not a usable memory chip, and the follow-up (stitching them into a real array and wiring to a PC) is where the interesting scaling problems will appear. But as a window into what semiconductor fabrication actually is — photolithography, doping, oxidation, metal deposition, all the steps the industry hides behind billion-dollar fabs — this is about as clear as it gets. The shed-clean-room framing also sneaks in a quiet point: most of this was possible decades ago. We just kept pushing the node size down until only three companies in the world could afford to keep up.