Making Ram At Home

RAM prices have been insane. Driven by the rise in AI, it is causing massive disruption to the GPU, cell phone, PC markets, and more. And a key reason it’s been so bad is that you have just three companies, Micron, Samsung, and SKH Highix controlling the industry where new supply is not something that can appear overnight. Building new fabs can take billions of dollars and years of work. So rather than wait around, I turned a shed in my backyard into a class 100 clean room and build my own semiconductor fabrication tools from scratch. Now the question is, can I actually make my own RAM? To start, how does RAM actually work? If we take a look and we zoomed in to one of the chips on a stick of RAM, what we would find are tens of thousands of rows and columns that make up an array where each intersection has a transistor and a charge storage capacitor. The transistor acts as a switch while the capacitor acts as a battery. When I turn on the transistor, it switches on the current and charges up the capacitor, storing one bit of information. I can then turn it off and hold the charge there. If I want to then read the data, I switch on the transistor again and the charge now flows backwards out so it could be detected. Though since that drains the capacitor, we have to refresh it and charge it back up periodically. But to build anything, we first need a design. I came up with a simple layout to have a 5x4 array that I could later stitch together where each intersection has a transistor and a capacitor. I’m aiming for a small transistor, a gate length a bit less than one micron when finalized. Each color here in this figure represents a different layer, like a different floor in a house. As these devices are made layer by layer by layer, kind of like a sandwich, silicon is the starting material for making RAM where here I take a whole wafer and then cleave it down to smaller workable chips where with a diamond scribe silicon nicely cleaves along certain crystal planes kind of like a like a box with a bunch of 90° angles. Now, after cleaving, there may be some debris on the wafer. So, I use a solvent cleaning process with acetone and isopropanol alcohol. This is meant to remove some of the particles off of the surface, as well as being able to dissolve any organics that may exist there. It doesn’t have to be perfect because later on on our next step, what we’re going to do is we’re going to convert the surface from silicon to glass. Anyways, the silicon chips are next loaded up into a furnace and heated to 1,100° C or 2,000° F. We’re talking volcanic lava temperatures. And this is in order to rust the silicon, effectively growing 3,300 angstroms of oxide or glass on the surface in order to mask and protect it, giving it this nice lime green color. With a layer of glass now on our silicon chips, it can be a little tricky in order to coat different materials. So I first apply liftoff resist. I coat that onto the surface. Now, this is normally meant, hence the name, for lifting off metal layers, but I find it actually works fantastic as an adhesion layer. This is then baked at 170° C for 5 minutes. Now, photo resist, our photosensitive patternable film, can be easily spincoated and spun onto this adhesion layer. After coating, it’s baked at 100° C for 2 minutes to remove some of the excess solvent. The photoresist uniformly forms a thin film of the silicon chips which is a little bit thicker than 1 micron. Photoresist acts as a patternable masking layer. UV light is used in combination with the mask where the mask only allows light to pass through designed openings and shine onto the photo resist exposing it. The 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. Basically wherever you shine light photoresist is now removed. So you have photoresist everywhere except for where the UV light illuminated. We use our first level from our design where the mask blocks light everywhere except for where our design is. The microscope stepper system has objectives which then shrink this pattern down right to where our sample is going to be. The higher magnifications further shrink that design down until you can’t make out the submicron features. We can see the design through the eye pieces and the camera where custom software controls the focus and exposure. The areas exposed to UV light are developed with our robot friend for better uniformity and ultimately removed forming our patterned photoresist. Using the photoresist layer as a mask, the patterned areas are dry etched where we selectively remove that glass layer to get back down to the silicon surface. With the edging step complete, we no longer need the photoresist mask. So, the photoresist is stripped in heated DMSO. Weird enough, people actually take this as a form of uh medication, but the semiconductor industry actually uses it as a strong solvent to remove photoresistant Lor. So, this leaves us with windows now in our 3,300 angstroms of oxide or glass on silicon. We did this as the oxide itself acts now as a high temperature mask for our next step which it uh involves forming the source and drain of our transistor. You can think of the source and drain as effectively the input and output terminals for our switch where the gate or switchy part will be later formed in the middle. We accomplish this by introducing dopants into the material in these windows here and that makes these regions highly conductive input outputs and this is by adding phosphorus into the silicon. So how in the world do we do that? How do we add phosphorus into the silicon? So there are a number of commercial products that do this. The industry also makes use of ion implantation, but that is ridiculously expensive and uh it’s too large of a footprint for a shed. So, Projects in Flight actually made a fantastic video on this, making phosphorus dope spin on glass as opposed to buying it. I tested the solution first on a test piece where the initial wafer is so resistive that a multimeter can’t easily detect continuity. But on a treated sample, it is now highly conductive and uh kind of working towards degeneratively doped, a very high level of doping, which is perfect. Repeating this on our main chips, we can coat the phosphorus doped spin on glass. And then we can proceed to bake it where the temperature is ramped up slowly in order to drive out all the solvents, but not cause undue stress or cracking. It looks good. However, there were a handful of glass precipitates that can form during synthesis. While this shouldn’t be an issue and it’s mostly cosmetic, uh, next time the better thing to do is to filter it in order to remove any glass precipitates. I made a calculator in order to model effectively the depth of our phosphorus dopen profile. What’s that going to look like? because what we want is that we want effectively a a more shallow uh profile for what we’re aiming for. And to get that, what we’re going to do is we’re first going to anneal with the phosphoros dope spin- on glass at 1,100° C for 5 minutes. Strip the spin- on glass and hydrofuloric acid. and then do a drive-in anneal at 1,000° C for 10 minutes. With the source and drain, the input output electrical terminals for our switch for our transistor. We can now focus on the switch aspect uh itself, the gate region of our transistor. Liftoff resist or makeshift adhesion layer is then coated because we still have glass on the surface. uh followed by our photo resist layer. Again, that patternable layer. We coat both of them. Again, with these layers down, we can then pattern and form the channel region of our transistor. To do that, we need this layer to be aligned between our previously formed source and drain regions. While we’re at it, we’re also going to align and expose the area that will form our charge storage capacitor, which is right above our transistors. After development, an HF is used to remove the middle oxide between the source and drain, as well as next to the transistor for our charge storage capacitor. This is because the oxide is too thick there and we want a custom oxide thickness for our gate and capacitor. The channel region of our transistor, the thing that does our switching, is the most critical region. As such, I perform a clean known as a piranha clean, which will viciously attack any organics and most metals on the surface. They call it piranha as it can eat material and flesh just like a piranha fish in the Amazon going after a chicken leg. The samples are loaded back into the furnace to grow the gate and capacitor oxides. We want this thinner so we have more capacitance and better gate control. So, we’re going to grow at a lower temperature of only 950° C. And if we put it in there for 38 minutes, we get 200 angstroms or 20 nanome of oxide. This is what the structure looks like now where we have our gate and capacitor oxides grown again at 20 nanome where we still have thicker oxide outside of the device. But because everything is covered in oxide, what we need to do is we need to punch some holes selectively through that oxide in order to later form electrical contacts to the device. More Lor and photoresist are coated and baked and then the contact cut mask is carefully aligned and exposed where after development we’re left with some small openings in the L and photoresist. HF or hydrofuloric acid is again utilized where that can go into the photoresist and Lor openings in order to remove any glass on the silicon surface thereby forming a nice pathway for electrical connections. For the final level in our device, we need to deposit metal to form the gate of our transistor electrical contacts as well as the capacitor itself. To do this, L and photo resist are coated and baked. And then we can align and expose the final mask level. Though this level is a bit different than our previous levels where our previous levels were all focused on removing material. With this, what we’re doing now is that we’re using the photoresist openings that we have in order to treat them as a stencil combined with a metal deposition process. A good analogy for this is that if we wanted to paint something uh such as a handicap sign over here, what we might do is that we might use a stencil uh in this case a metal stencil where we could deposit material by spray painting for instance and then when we remove our stencil, we’re just left with the design where we want. Well, that’s effectively what we’re doing here at the nano scale with the pattern that we’ve created. But instead of spray painting, what we’re actually doing is that we’re effectively spraying the surface with metal atoms, in this case aluminum here, where the sample is loaded up into my spotter system where we have argon that hits the metal target and then the metal atoms fall off and are effectively sprayed, so to speak, onto our sample. So, as you can see, the metal has been uniformly coated over the sample, uh, except where there was some kept on tape at the edges there. So, how do we effectively remove our stencil? How do we remove the photoresist layers here? Well, we can stick that in once again DMSO and heat that up. It’s always cool to see where we can start to see the metal buckle and then eventually start to peel off of the surface. If I apply a little bit of agitation there, we can aid in its removal just leaving the pattern that we want. Looking at it under the microscope post lift off process, what we can see is that we can see our complete structure. Now we have the transistors, the capacitors, all the connections that we need in order to have our DRM array. If we were to look at a cross-section, this is the structure that we have. And then this is how it matches up to our earlier little cartoon where we can have current flow and be controlled by the transistor by our switch where it can charge up our charge storage capacitor in order to hold a bit of data. Now, as if the shed wasn’t enough, inside in my living room, I hold a whole suite of test equipment. So, I got my little lab assistant here, and I’m going to fire up one of these semiconductor parameter analyzers in order to look at the results. Because these devices are, of course, at the nano scale, you can’t just attach regular wires easily to them. So, for testing purposes, what I have set up here is that I have a number of these micro manipulators with some incredibly fine probe tips in order to feed electrical current and voltage into the device and then be able to read that out. When testing the transistor or switch, I get the following results. Each line represents a different gate voltage where I can have different levels of current or nearly no current depending on the gate voltage I apply. This acts like our switch or more specifically it can act like one of those dimmer switches where as I turn and adjust I can tune the level of output power. But for RAM we could just settle for basic onoff operation. Normally though in a transistor current saturates. These lines stay flat. It doesn’t shoot up at higher voltages. What we’re seeing is a short channel effect known as punch through. So because the source and drain are less than one micron away from each other. As we increase the voltage that we apply, they can effectively merge. And this leads to an increase in current and a loss of gate control. Now this is fine as long as we operate our devices at lower voltages but it certainly shows the challenges of scaling. So let’s look at the other half of the equation the charge storage capacitor that we have. So here I have a CV plotter which is a very sensitive instrument for reading up the capacitance as I scan through and then vary the voltage across the capacitor because the capacitor will have a minimum and a maximum capacitance of course in which it can store and that maximum capacitance I recorded as 12.3 poparads which is pretty close to the perfect ideal theoretical of a little bit less than poparads which I had designed for. Looking at them together now as one individual DRAM cell, the transistor can very rapidly and quickly charge up our charge storage capacitor to 3 volts uh within a couple hundred nanconds, which is great. And then it’ll hold that voltage on there, but it’ll slightly bleed off over time where we can only hold that charge on there for a little bit over 2 milliseconds before it’s completely done for. We need to charge it back up, if not sooner. Commercial DAM can hold its charge for greater than 64 milliseconds. So, in this design, it needs to be refreshed at a higher frequency. Okay, this is awesome. First time ever, RAM has been made at home. But while you can store data on it, you can’t run Doom on it quite yet. This is just a few cells to prove it can work. The next step is to take these cells, stitch them together in order to make a much larger array. Then we can hook it up to a PC. Stay tuned. Massive thanks to everyone subscribed on Patreon who helped support this work. Link below.