Resistive random access memory {ReRAM} Explained {Computer Wednesday}

Added: 2026-05-31

[00:00:00]Hello YouTube viewers, welcome to my channel Science Technology. In today's show, computer witness t we're going to talk about resistive random access memory or ream. So let's dive deep into it. So what exactly is the problem right now? Problem is that world is hunger for more compute and primary component of that compute because be mindful your CPU or your GPU could be only as fast as the slowest link. Right now RAM is the slow link. We want faster and efficient RAM because if you keep increasing the amount of RAM, power consumption on servers could go stupid and GPU stupid meaning it will be just too impossible to cool. So uh fundamentally we want efficient RAM and we want faster RAM and there are fundamental core limitation of the technologies that we are using. This is one thing you have to understand. If you understand the core science of something there is a theoretical limit all technology like for example can you ever make a CRT as energy efficient as OLED no like under known laws of physics that will simply not happen that's a physics level system for example like if you're burning uh let's just say oxygen and kerosene how efficient you can make a rocket engine in vacuum versus uh oxygen and hydrogen oxygen and hydrogen would be always better why because oxygen is the lightest uh sorry hydrogen is the lightest element it will give you higher specific impulse. These are physics level thing. So in computer electronics there is a physical level thing. Many times we feel like oh technology improvement trust me when uh people figured out how to make silicon based system they calculated what is the theoretical maximum and we have not reached it like you see Mo's law but here still we knew about that mathematically we knew. So fundamentally how we are using RAM it's super fast but it's volatile AF as in the moment power goes delete done go home that's why your laptop sleep requires electricity otherwise if you open it up afterwards it's like I have no idea who are you I have to let do everything again so that's there and NAND is nonvolatile good but if you compare that NAND technology to this uh basically floating gate technology to RAM chips it's like bro it's not even in the same league and again this on a fundamental engineering level how we are building things it cannot be bypassed and yeah you can compare like okay 1980s RAM chip to like you know 2026 flash chip but that's cheating assuming within the same technological bracket the output would always be the same RAM will always be faster but again it will always be volatile and NAND while it will be slower it will be nonvolatile and our AI needs bit too much like faster faster and faster system. So, but wouldn't it be great if we could store multiple bit per cell of a RAM? We could do that right now with NAND systems, but we can't do that with RAM system. So, I'm going to go in each tech layer and let's understand things.

[00:02:52]So, tech layer number one, which is RAM.

[00:02:54]Now, RAM is the core of modern systems.

[00:02:56]And when I say core, I mean fundamentally your CPU's first task is initialize the goddamn RAM. That's why if RAM has a issue, your system does not go. It's like it's off. It's like dude I have no idea what's going on. Like if RAM initialization fails, it's like it would be very difficult for CPU engineers to figure out enough memory or enough compute to work without RAM.

[00:03:16]Whereas like even just to throw up a error message, it's like generally error message comes from your motherboard. If RAM is not working, the blinking light is RAM. So RAM is core of it like core core. Now why the heck it's so important? Well, here's the deal. It has super easy access to each cell, meaning each of the uh bytes. so to say you can if you want if CPU wants it it can go to each bit basically let's just say your computer required a data right it puts it on a table and because it's on a table it has the luxury of seeing each lines of it like think in this way like you have multiple open pages of multiple books on the table is the RAM library is your hard drives or your SSD so when you are something everything is open like that you can write here here so same thing which row which column tada that exact bit I need to change this one bit CPU can count on RAM to A find that bit B change that bit all of that will happen seamlessly reliably and rapidly so this is why we can read and write data to RAM very fast because each bit is accessible it does not need to take out the whole bite change it then put it back in no everything happens instantaneously almost so how the heck we are storing it well if If you have electron microscope zoom into these puppies, they are look like something like that. So what is those blocks?

[00:04:39]Capacitors. They're holding charge. They are capacitors. Now each uh benefit of having each so much wiring for each cell uh is that you can address each and single cell is bite is available. Like generally computer ask for a whole bit uh sorry bite whole bite block rather than each bit. But if it wants to it can do that. We don't do that anymore. It's like you have enough throughput. And by the way nowadays CPU itself has multiple level of cache that's what it has.

[00:05:09]So fundamentally and when you are talking about capacitor when you're making capacitor this tiny this self discharge meaning you have capacitor cool you put voltage that's let's just say you're on state great uh you're reading it great over few minutes dissipates. So what do we do? We keep refreshing it. Meaning computer reads it. Oh, there is a charge. Yes or no?

[00:05:29]There is charge. Keep charging it again.

[00:05:31]Uh there is no charge. Don't charge it.

[00:05:32]So again, so they want to maintain the state of it. You do not want to rewrite everything on this uh because of refresh. So refresh has to do two tasks.

[00:05:40]Not only it has to check every cell and then it has to reflash everything and it happens almost every time. That's why the moment power goes away the end all of the data in this goes away. Now to be fair, if you dunk these puppies into liquid nitrogen, uh the moment you turn it the power off, it should last long enough. Should last long enough. Some hackers did use that method like you know liquid nitrogen, it did make it froze enough and for some reason electron did not leak and voila you can read it. So fundamentally it is power hungry. That's why your laptop's sleep state does require power. It is trying to maintain the RAM module meaning it has to keep uh charging all of those capacitor fundamentally and because of so much wiring and the design and the architecture of these puppies uh they are fundamentally low density packing meaning this big of a board will have much smaller capacity compared to your M.2 to SSD. So that's on a physical engineering level. Benefit hyper fast.

[00:06:36]Each bit is accessible. Each bite is randomly accessible. But penalty super duper low density and power hungry as hell. Like you might think of RAM, does it really consume that all? Look into servers and you're like, yep, it does.

[00:06:51]Like bonkers amount of power.

[00:06:54]Now we come to NAND. Now NAND is nonvolatile. That's the great part. Uh you can turn it on or off. It does not care. It will store the memory. It's what we call SSD. Uh but it's where the heck it's storing charge. It's storing charge like a capacitor but not in a capacitor. It's storing it in gates. Uh basically if you know MOSFET something similar to that. So there is floating gate and it's in that gate it's charging. It's like so if the gate is empty like inside the gate there is no charge that's zero and it can put a charge into it that's one. So it can store that and uh based on your system so to say how much charge is there you can read out whether it's zero or off.

[00:07:33]Now be mindful people figured out very early on it's like hey why only have on or off? What if you have multiple levels on it? So you could do that. That's what we call multi-level cell uh triple level cell or quad level cell. So if you have nothing like nothing zero shun that's 1 one. If you have sum that's 1 0. If you have bit more that's 01. If you have a lot more that's 0 0. Now you're like wait a minute that would require a very sensitive circuit. Yes that's why nan chips on principle can work at very high voltage. Like internal voltage could be as high as like 10 20 volts. Like wait a minute why? Well it the higher the voltage the easier and cleaner the signal becomes. If you try to distinguish each levels inside one volt.

[00:08:16]Yeah. No noise background noise alone will make it very difficult. So that is the fundamental uh limitation. So and to increase the packing efficiency all of these is in shared bus you see each is a P channel and each of them are N channel this is how it's physically constructed so you can't access each one of the nodes specifically for right operation read operation you can manage if you want to do a read operation that where it almost behaves like RAM it's almost to do almost not as fast but almost but right yeah it has to flush the whole thing and then write it and there lies the core problem that's why uh write operations on SSD is always slower than the read speed. It could only be compensated by one way. You over provision it so much that bus becomes a limitation. For example, it's very easy to do nowadays with SATA 3 interface.

[00:09:03]Like you have a SATA SSD. It can be made so fast that right speed is higher than uh let's just say 600 650 your SATA controller can't do faster than that. So it's like bro it's saturated on theoretical read would be like 1200 but here the the link speed is 600. So write speed will always be slower because it has to take all of the data. Put it into a RAM. That's why your SSD has a RAM.

[00:09:26]Put all of data into the RAM. Flush this whole thing. Rewrite with a new data change again. And there lies the core problem. So and because you are dealing with charge not a capacitor and again it could have multiple levels. It requires time to stabilize. fundamentally on a physics level all the analog to digital converter that's actually doing the measuring it's like yes no it's not like it's not as fast as it fundamentally requires charge to travel to equalize to stabilize it otherwise it will not work now it's getting faster but physics limitation it will never reach the speed of RAM like it's on electron level you can't do too much to it it's too slow for RAM use and too leaky for long-term data archive even though it's nonvatile is not classified as a long-term archival meaning if you are a corporation you will not store your tax record of 10 years into this you will store it on a magnetic tip and I hope you keep replacing it regularly so fundamentally it does work and as we have um improved the technology we have two choices for example if you are running mission critical system you will buy a single level why the voltage would be huge as in like oh if it's 2.5 volt it's on if it's anything above 1.4 4 volt is still on. Anything below that is zero. So that's your gradient and you have room. So even a degraded cell works and you're like okay there is no moving part but what's the degradation? Yeah that electron barrier that basically dilectric you have you forcing electron through it and pulling it out. Every read and write operation read does not do damage to it. Write does. So every time you are writing to it you are eroding it. And as in like take this uh chip fresh out of the factory, slice it, put it in electron microscope. The dialectric layer would be like GG pristine strong. Use it like for 3 4 years in a heavy compute environment. It will be destroyed. So it's a physical wear item like that. It's physically that's another aspect of it. RAM will work for decades. This will not physically not that electron layer will be like bro I don't. And again uh to compensate for that degradation either a have large enough voltage range like single level cell whereas like bro I have so much wiggle room that sloppiness of that layer will not compromise too much but uh you want data high data package you could have quad layer which translates to 16 voltage levels. Yeah that's why quad layers are not used for mission critical things. It's great for packing, meaning your micro SD card could be two terabytes, but not great for longevity. Meaning you you should not trust that puppy for long term because any failure would be such a big failure that it will corrupt large segments in one go. That's why like servers will uh if it's mission critical, it's a single level like they're not even talking about anything else. It's it better be a single level.

[00:12:10]So too slow for RAM, too leaky for long-term data archive. Meaning even if the barrier layer is not degraded it's still a kind of charge electron kind of stuff it will slowly leak through not in one year let people do expect like 5 to 10 years bit rot should start to happen 10 20 years good luck so there is fundamental it's a good thing off uh it works when it's off great it uh has very high packing density great it allows you to have each cell uh have multiple state store that's great but does have some wear limitation and again this is another thing if you use SSD as a RAM even for your old computer even though speed will match you will find that it will not last very long it will degrade and again people are like when they used micro SD card for Raspberry Pi as a operating system disc used to wear down very quickly they were TL TLC basically micro SD cards they are not classified as high priority so rarely they are single level so they degrade very quickly Now we come to the RAM. Now you must have understood all the concepts of the previous one. Then it's just that what if you have a cell instead of storing charge what if you are storing a resistor. It's like wait a minute what yeah what if it's a ste it's almost like a potentiometer and you're setting its state. So it's much faster to read compared to NAND simply because it's much quicker to read a resistor than it is to like okay how much charge is there? And again because you are talking about gates, you are talking about floating koms and all that you do need some time to equilibri. That's on a fundamental physics electron level. Here it's like super fast and resistance are long-term stable. There is no energy into the system. So it is like bro I'm set like if you set it low resistance state it will remain low energy state. There is nothing to leak out once you change it.

[00:14:03]It's like there. And if you want to rewrite it you uh you know increase it with a voltage shock. so to say. So residence is long-term archival again this is theoretical at this point in time that it could work for 20 30 years without bit rot that's hope and again on principle it should happen because uh you know our nan chip does store memory it's just not great at it but does store memory so this should be exponentially better because there is no high energy state that wants to leak out it's once you change it it's stabilized and it can have multiple state per cell each cell could be like hey low resistance Say high resistance stage one, high resistance stage two, high resistance number three. So each of them could have multiple gradient and be mindful the gradient. How do you measure the gradient? With a resist meter. So it could go from like you know 1 megga ohm.

[00:14:51]That would be like you know fully fully open that would be 1 mga and then you can keep dropping it down to 10 kiloohm.

[00:14:58]That's a lot of room. That's like going from here to here. That's a lot of room.

[00:15:02]And meaning in principle I can see that one cell could even go up to 256 uh levels and should be able to store one bite. Each cell if it can store one bite of data that's like shut up take money that's like you have no idea what level of technology we are talking about if that can be attained and have going from 10 kiloohm to 1 megga you have enough room where you can slice each of them with enough margin that there will not be noise in the system. So that's the exciting part for people. We are like bro we're going to start with obviously on and off just to test everything and validate but quickly switch to like you know 16 17 uh 14 states day one like that's everybody who's working on it is like yeah we're going to do that. It's like it's we're not even trying to make RAM like RAM. We are making something that will be from day one much higher density than RAM and very quickly like two three uh iterations afterwards it will be like your SSD on your RAM stick.

[00:16:01]So and each cell is addressable. The geometry, the designing, the engineering is done in such a way that it acts like a uh ramp. Basically rows and columns are fully accessible. It's not like whole blockade.

[00:16:14]So you have that and it does not require refreshing like a capacitor does.

[00:16:18]Capacitor requires recharging. This is like bro once you set it forget it.

[00:16:21]Meaning once you close your laptop let's assume the battery forgot like you you lost your battery. The it was completely discharged to 1 month and then you plug it again. You open it up, it's like bro I got you. Your mouse pointer was here.

[00:16:32]It will be here if you have set it in that way. So fundamentally and how do you construct it? It's a electrode and bottom electrode and you have metal oxide with a conductive filament. Now this filament is the black magic of it.

[00:16:46]Uh it's basically electron mobility kind of thing happening. Uh again mo it's like whole pairs and all that. So that's why it's changing resistance. It's not changing like oh it's now a cell. It gives you voltage now is just changing the resistance and to read it again resistor uh whenever you have like you know multimeter and you're trying to read it generally sends uh some uh current through some amount of voltage through probes. Uh so again if you are using the probes how the heck you going to make sure that you are not damaging it or changing it? Well for reading the value you're going to use low volt as in like give or take 500 m volts. Uh for writing it you could be as high as 5 volts. So writing is a very strong process and reading is a very low energy process. So this actually has a advantage that if you want certain sections to be not you know overwritten it saves energy. It's almost like OLED where black is off is the same thing here. It's like hey I don't have to spend energy here simply because reading it I can read it at 500 molts. I don't even need to spool up my DCDC converter and pump up to 5 volts.

[00:17:49]So that's the whole point. Uh top electrode, bottom electrode. Now I cannot talk about the chemistry here simply because a I don't understand it.

[00:17:56]B it's not sorted. Every research paper I'm going through is different. Some is like titanium oxide is the one thing that I can almost guarantee that somebody will bring out unless we find something better somebody's going to bring out titanium oxide system. This whole system so this is the readm each cell basically the structure is designed in such a way that is RAM. The packing efficiency is much higher. It does not require refresh and reading it is a very low energy process and it can retain its state even when it's off and on paper on principle this puppy should outlast a NAND flash chips in terms of memory storage in long term.

[00:18:32]So what can we expect in the future? Now here's uh this could be the thing as in this is what we classify as paradigm shift like if you uh studied computers in 1980s specifically 70 period like we barely had 4bit computer at the starting of 70 and the end of uh uh '7s we ended up with um inter x86 so to say. So that was paradigm shift. has the potential of that paradigm shift because if done correctly if what people are expecting can be done. Let's just imagine a world where from engineering point they can't pack like again they can't pack too many compared to a nan chip so they can't beat a nan chip in that but because they have resistance and they have the luxury of having so much dynamic range in the resistor. What if each cell is a whole bite? It's like yeah you can access it but it's the whole goddamn bite. So that on paper could create a system where you do not need any SSD or RAM that you just have reams. The end it is the memory the end computer will behave as in like whole of its operating system is in the goddamn RAM like there is no slow storage there is no fast storage. So that's from a uh like you know paradigm shift point of view. Second, why people are pouring so much money into it right now is that it can act like a analog computer. For example, uh whenever you hear about uh AI, what is AI? Mostly it's a neural network nodes. What is that node? You send a one data, multiple nodes work on it and you have output.

[00:20:01]Here's deal. You can do that with resistors. Again, those who work in analog domain, it's witness day for them. But for digital people, it's like what? Uh so imagine you sent 5 volts.

[00:20:11]Okay, how much voltage should be on the output end? that can be defined by the resistors. So you have multiple resistors, right? Each node have multiple uh vectors so to say all those are arranged in series. So your the collection current will tell you what happened like I send this data it goes through this much what is the output.

[00:20:29]RAM is acting like a computer. So you can do AI summing operation on a hardware level fundamental hardware level. Uh although let all the engineers who are working on it they are very clear this is not for training data.

[00:20:41]This will not work for training. It assumes you know the training like you know the weights and values then you want to uh project that weights and value into these hardware and this hardware will be like exponentially fast meaning you will take a data center to train it and that final AI could run into something like this without consuming energy and you will not be like oh it's a bit slow compromise no it will be as fast again assuming the training went well training to this projection can be done in such a way that it's like bro I got this it could take megawatt of power to train AI and it could take microwatts of power to run it on a hardware level. It's a hardware domain thing. So that's a very interesting that's why people are pouring money into it and they have like haphneium, tatt, tantelum, zaconium, this and that. It's like that's the only thing I know is titanium oxide. Uh so but again some people are also working with the leader but you get the point.

[00:21:31]So it does need time to cook if you try to look into its uh source materials some of them are ancient like not old ancient. So it does need time to cook and again we do not know for certain whether this will become the thing in the future or not. It could, it could or it may not.

[00:21:53]So this was my presentation on basically ream. Hopefully you have liked it, learned from it. In that case please click the like button, share it amongst your friend. That will help me a lot. If you enjoyed it, please click the like button. If you didn't like it, didn't enjoy it, I urge you to press dislike.

[00:22:05]Press it twice to show me excess permit.

[00:22:06]Please leave a comment because I do try to reply to all of them. Subscribe.

[00:22:09]Press the bell for free.