This week’s article is a bit all over the place, but it ended up being a lot of fun: we’re going to dive into the world of how chips are manufactured (the electronic kind) and dive into the theory around making one of the world’s most advanced storage devices.

Originally, this article was going to be on topology. If you follow a lot of Arxiv posts, the term “topology” often gets thrown out as a sort of academic buzzword. In the long run I’ve been wanting to make a series on these buzzwords (technically the article on Markov Chains was the first) so topology was naturally my next target.

Alas, the term “topology” had a lot more applications than I expected. I knew it was a field in mathematics, much like algebra and calculus, but I figured if I was going after a use case the specific form of topology I needed would be easy enough to track down. As it turns out, topology is popular! Pretty much every trending field – LLMs, quantitative finance, self-driving, robotics, electronics, etc. – uses topology in one way or another. Unfortunately for me, it’s never in the same way.

While I intend to cover other use cases in the future, one in particular caught my eye – topological insulators, or TIs. The term refers to, in a simple sense, materials that allow electrons to only flow on their surface. Usually, if you zap a material with electricity, it uses the x, y, and z axes – in other words, the electrons will dive deeper into the material, and basically become useless. This is not the case with TIs, thus you only have to worry about the x and y axes, and can use all sorts of fun surface topology tricks to maximize your surface area and get the most bang for your buck in terms of circuitry.

And while I could’ve looked deeper into what these surface topology tricks were, I felt myself getting pulled more in the direction of these insulators. My electric engineering knowledge is pretty lackluster, and especially my knowledge of materials. How exactly do you make electronics with TIs?

I had an idea: let’s learn by making one on our own!

How Hard Drives Work

Now, full disclosure: in order to manufacture a real hard drive using topological insulators, it requires equipment that’s somewhere in the range of ten million dollars. And unfortunately I left my ten million dollars at home when I was writing this, so the equipment was completely out of my hands. I’m also not Sam Zeloof, which should really go without saying.

All this is to say that our manufacturing of this hard drive is staying on the whiteboard. Still, it should be enough to get something out of it.

I also chose a hard drive rather than something like a CPU or RAM because a hard drive, at a fundamental level, is very easy to understand. Literally all we’re doing is writing data to the system, reading it, and deleting it if necessary. TIs can be useful with CPUs and RAM as well – the Microsoft Majorana quantum computer is named after a particle found in TIs! – but we’re already biting off almost more than we can chew.

A drive consists of two things: a controller chip, and the storage itself. The controller chip really just plays the role of translator between the storage drive and the computer – if you save something as a .jpeg, it is the controller chip’s job to say “Hey, Windows! That’s the jpeg, right over there!”. The chip is made using just standard electronics, and so we aren’t going to concern ourselves with it too much here.

Now, an important thing to keep in mind before we get too deep into the weeds: the way TI storage devices work is fundamentally different to how old storage devices work. It’s part of the reason why storage made using these TI materials are often called “next-gen storage” (anyone getting PTSD flashbacks to the cybersecurity article?).

In the old times of computer storage, we used magnets. These drives (known as HDDs, where “hard drives” get their name from) would store data by magnetising chunks of a spinning disc to either a 0 or 1 in binary. Of course there were unfathomable amounts of these chunks, which is why we can save data on them as large as a terabyte.

However, even back then we knew magnets were not a great permanent solution. The first problem with HDDs was that, especially in its infancy, these spinning discs were loud. Think 2x louder than your GPU fans today. And even when the discs were able to get quiet enough, there was still the issue of read/write speed – HDDs are miserably slow to load in data, especially for something intensive like a 3D video game or a commercial database.

The long-term play – which eventually came into fruition in the mid-2010s – was to move to solid-state devices, or SSDs. These got rid of magnets entirely, and instead focused on electrical charge. Now, there was an array of junctions, known as floating-gate transistors, which trapped electrons in order to mark them as either a 0 or a 1. This process was significantly faster than using the magnet method, and is the de-facto standard today.

But… could we get better? SSDs were still not optimally fast (it took awhile to capture an electron) and they were more likely to suffer from degradation (the electrons eventually escape after 5-10 years and the data becomes corrupted). This brings us to the latest idea, and the one where TIs really shine: using an idea called spintronics.

Instead of using electron charge, we now use electron spin. Without going too deep in the weeds here – if we’re calculating based on electron spin, all we need is for the electron to pass by the junction, not get captured by it. Whatever the spin of the electron, it basically affects our junction in such a way that it is either a positive spin (1) or a negative spin (0). Same effect, now faster than ever! In addition, we no longer need to hold any electrons, nor rely on a slowly decaying magnet, which means our data is held (technically) forever. You can start to understand why people call this “next-gen”.

While we thought of the idea of using spintronics for a while, before it wasn’t as practical due to how we’d need to place the junctions. If you want to capture an electron, it doesn’t matter if it’s deeper in the insulation – it will get attracted to its correct location, and be trapped that way. But if we want to use spin, we need it to pass through the junction by itself. In normal insulators, the decay we receive from electrons falling deeper into the material makes this pretty much unusable. But for topological insulators – those ones that don’t have electrons go anywhere but the surface – it becomes much easier.

Alright then. Now that we understand what makes our hard drive so super, how do we go about making it?

The Layers of Our Hard Drive

Our Super Hard Drive (as I am calling it, because I loathe the name next-gen) consists of the following layers:

• TOP, Read/Write Circuits
• Magnetic Junctions (Cobalt)
• Insulating Barriers (Magnesium Oxide)
• Topological Insulator (Bismuth Telluride)
• Grounding (Platinum)
• BOTTOM, Device Substrate (Silicon)

The best way to think about this is like a highway. Our cars (aka electrons) drive on a highway (our TI, most commonly Bismuth Telluride). There are guard rails that stop our cars from crashing (Magnesium Oxide) and clearly marked road-signs telling them where to go (the read/write circuits, made using standard circuit materials outside of our scope). At some point, they pass a toll gate (the cobalt junction) which marks that they were there (1 or 0). Then, finally, they exit the highway and continue down their own path (the platinum).

Hopefully this analogy, combined with the details I gave you earlier, makes this pretty clear. The fabrication process for this is pretty cool, and while I didn’t ask ChatGPT for much details since it was out of scope (I might make a whole different article on fabrication) This is the general summary it gave on how we’d construct it:

1. I didn’t mention the silicon in my highway analogy. It’s because that’s not related to the drive itself – it’s actually what we print the materials on! Think of the silicon like paper, and the materials like different types of ink. We start with Bismuth Telluride, where we vaporize Bismuth and Telluride separately, and let it slowly deposit onto the surface of the silicon inside of a vacuum chamber.
2. A thin Magnesium Oxide layer is deposited on top next, using a method called Atomic Layer Deposition. Fuck if I know what that means.
3. Using something called Electron Beam Lithography, we etch tiny holes in the Magnesium Oxide layer where the magnetic junctions will go. Then we vaporize Cobalt and let it form in these little pockets. This stuff is all like, incredibly precise by the way. If it is not exactly 10 nanometers thick, the entire drive will not work. You see why this requires its own factory?
4. Finally, we can take the silicon off (the new drive should be thick enough) and replace it with our grounding layer of Platinum. On top of the Magnesium Oxide, we etch some more stuff – this time it’s the circuits, or roads, that we let the electrons take.
5. We package the whole thing in an SSD-like enclosure, with a controller chip attached. Now if we plug this thing into a computer, the computer will know to send zaps of electricity into our new super drive, which will write or read data with the electrons depending on what is needed.

So… Why Don’t They Exist?

You remember when I told you that you’d need tens of millions of dollars of equipment to do this? Did I mention it was tens of millions of dollars in custom equipment?

Let me put it this way: if you build a fabrication plant to make SSDs, you can only make SSDs. The fabrication’s abilities do not translate to CPUs, GPUs, HDDs, super drives, or what have you. This is in part due to that “ultra super specific” requirement I mentioned during the fab process.

So, if you want to make super drives, you have to invest millions more dollars into brand new plants. This is not usually something that companies do unless they know for sure that there is a demand for this sort of thing. And even if there is a demand for it, the company still needs to make back its losses somehow – so it translates that cost to the consumer. That’s why brand-new hardware usually costs 10x its first few years before companies eventually get into the groove and build enough factories to take prices down.

So, it might be a while until we get super drives. Fortunately, the demand is there: spintronics is something that is necessary for quantum computing and advanced AI manufacturing, which is what is number one on the radar of pretty much every major R&D house out there. And now we have an advantage on the learning curve!

[Image omitted]