I will be talking about AI in a little bit broader sense, and especially how ReRAM is really a great solution for AI, whether it's edge devices or in-memory compute, etc.
So, well, let's dive in and go to the first slide, the next one. So, the first thing is AI, and I'm sure this was discussed in previous presentations, and I apologize I wasn't there. It really consists of two parts: there's the training and there's the inference. And we're used to talking about the training because for a long time this was the main activity in AI. You just need to train all of these models and learn. But at a certain point, you get to having these models, and now you train once, you infer all the time. And inference has been growing over time, and 2026 was actually kind of a pivotal year where for the first time inference really took the front stage.
Right now, this is the bigger spend already in AI, and it's really becoming the main activity going on both in data centers and in edge devices and everywhere. So if we go to the next slide, you know, let's look at the characteristics of training and inference. As long as we talked about training, you know, we were building models, we needed big data centers, everyone was talking about data centers, you know, maximum throughput, you needed a power station next to each data center, you know, huge power consumption.
And obviously, when you're building the models, you're constantly improving them. You're constantly updating them. So that was the main activity, and this was done in data centers. It was very focused on these specific facilities.
But now that the world is starting to focus more and more on the inference side, you're really now taking these models and deploying them. And you deploy them everywhere. You deploy them in the cloud, you deploy them in the edge devices, of course also in data centers, but in industrial and automotive, in everywhere. Now you're basically using the model. So you're reading it all the time. You need to read very fast. You need to be very energy efficient because it's everywhere and many battery-powered devices, etc. So it's really just, you know, billions of deployments in every place, and the characteristics are very different. So from talking all the time about power and capacity and, you know, just how fast you can compute, we're now focusing more and more on power consumption, being able to run with batteries, on speed, and basically we need to be close to the compute engine to make things work fast, and we're scaling down to the small geometries within friends. So if we go to the next slide, we basically, in addition to everything that I said, and this is something that is now growing very fast, we have the physical AI, the robots of all sorts that are coming into the picture. And these guys, not only do they do inference, they need to move around, they need to respond to a real-time environment. You know, when they're moving, they need to recognize things and react instantly. When they're picking up a delicate glass, you want them on the one hand to hold the glass so that they can pick it up, but not press too hard that they break it, etc. They really need to be able to react all the time to their surrounding. And that's really key. On the other hand, and these things have so many motors and so many, if you think about it, all these joints need to be moving, and you need to control all of these things, and each one of them you need to be really analyzing and reacting in real time. On the other hand, you don't want the batteries to run out quickly. You don't want this robot to be sitting and plugged to the wall most of the time. You want it to be moving around and doing its thing. So the requirements are even more extreme than just regular inference.
And if we move to the next slide, and we'll see, you know, there are two really main places which we're talking about. There is what's happening today with using emerging memories for AI, and this is mostly in the near-memory compute domain, and there's the longer-term in-memory compute. So I'll start with the near-memory compute, and we can go straight to the following slide and look at the demands from AI inference.
So I started talking about it. You need to have high density. You need to be able to have large memories, a large amount of data on the chip. You want it to be near the processing unit. You need to be reacting very quickly. It needs to be ultra-low power. Again, most of these things are battery-powered and robots and stuff. It needs to be very fast so that you can react to the real-time inference. Endurance. I mean, while you're reading most of the time, these models are updated every once in a while, and you still want to have high endurance and be able to make more updates over the years. And of course, everyone wants to go down to the small geometries. You need to be able to go down to the advanced process node, and eventually you want to go to in-memory compute. So these are the requirements, and ReRAM is really such a good solution for this. The ReRAM bits are four times smaller than SRAM, so you can really cram a lot of memory on the chip, which is very important. It's very low power consumption. It's very fast. The endurance is much better than flash, 10, 100 times better than flash, and we can scale down to the small geometries. So really, ReRAM is a very good answer to the needs of near-memory compute.
And if we go to the next slide, we'll see, you know, what's happening today. So as I said, inference is happening already, and right now in the majority of the cases, there still isn't ReRAM in those chips. What people are actually doing, and you can see this in the traditional AI, they have an AI chip which is obviously in the smaller geometries, 22 nanometers and below. Now they need the coefficients for AI, and those have to be in a non-volatile memory because when you turn the system on, they need to be there. So they have an external flash chip. Flash cannot go below 28 nanometers. So you have an external flash chip. When you turn the system on, you're basically copying everything from the flash into that SRAM, that big SRAM inside the AI chip. And then you can start running. So, you know, obviously having two chips is more expensive than having one. The boot time takes now very long because you need to copy everything into that SRAM, and that takes time, which by the way causes security risk. It's so easy to eavesdrop on that communication, so a lot of issues with that solution. But now that ReRAM is available, there's more and more interest, and people starting to go in this direction of replacing that SRAM with ReRAM. At that point, you have the non-volatile memory inside the chip. So you can throw away the second chip. When you turn the system on, it's immediately ready to run. So basically what we call instant on, you removed that security issue, and I already mentioned ReRAM bits are about 25% of the size of SRAM bits. So now you can either have a smaller chip, which again saves you money, or what most people would want to do is just have a lot more coefficients and have much more accuracy. So ReRAM already, just in the near-memory compute, which is something that is very easy to do today, it already gives real advantages. It's much more efficient, and this is already where the industry is starting to go to. But obviously, you know, this isn't the only thing. So let's go to the next slide.
And, you know, I mean, first of all, of course, ReRAM is available today, and we also have it qualified at DB High-Tech. It's qualified at SkyWater. It's very close to qualification, on the way to qualification with Onsemi, and of course we're working with Texas Instruments, and there are many others that we're engaged with and moving forward. So it's a proven technology. There's silicon out there. It's qualified, being licensed, and really we're seeing this ramp up in the near-memory compute. There's interest and a lot of work going on. So now moving forward and looking at where we're heading, which is in-memory compute, and we can go straight to the next slide.
The real issue when you're doing inference, when you're doing AI in general, and when you're doing any type of compute, is what's called, I mean, it's really the memory bottleneck here. Almost all devices that you know today, and not just AI, they are built in what's called a von Neumann architecture. They have a separate compute engine, a CPU or microcontroller or something, and a separate memory, and the data keeps going back and forth between them. Now, this always was a challenge in AI, which really works on all of this memory. By the way, yeah, when you're dealing with a regular compute engine, there's what's called the locality effect. So you're addressing the same area in the memory most of the time. But in AI, you're really going all over the place. So you can't even copy that local part into the chip and what's called a cache and try to be more efficient. You really are all the time accessing memory, and in current systems the statistics show, you know, up to 80% of the power and the time are wasted just on moving data around. Now imagine, and let's go to the next slide.
So in this traditional memory, as I said, you have that bottleneck, and imagine that you could now do the compute inside the memory. I mean, that would be unbelievable because you could save this 80%, you could do things in a much, much more efficient way.
And this is actually what happens when you're using ReRAM for in-memory compute. So we can go to the next slide. And basically what happens is that when you are doing in-memory compute, you are almost all the time doing vector-matrix multiplication. That's the activity that you're doing in the vast majority of the time. Now in that case, you can build the memory in a slightly different way than what you would for storage, and you can have the matrix in the memory, and you can have the vector written into the memory so that the multiplication is done on the spot without any data moving. And what we do, we leverage some basic electronics rules. Anyone who will learn the very basic electronics knows these rules. So what we do is the weights are actually stored in the ReRAM bit. Now the ReRAM bit is actually an analog memory. So it doesn't just store zero or one. You can actually play with the resistivity and go from low resistivity levels to high resistivity levels. So you can have different weights in the bits, and you can now play with the voltage as the input. You can use, when you're writing into the bit, you're actually applying voltages. And there's basic Ohm's law that says the current is the result of multiplying the voltage with the conductivity. So that is what you're doing here in step three. You can basically write that voltage or use that voltage as the input to that memory bit, and you can get the current coming out. And if you actually just connect all those currents together in a very simple way, you have the result of the multiplication. So this is how you do a very simple multiply-accumulate using a ReRAM, and it enables you to save, as I said, a very significant amount of time and power.
And if we go to the next slide, we can see the results. A US-based American company, they took our samples that we're using now to play with in-memory compute, and they ran their own tests, and you can see the results that they got. So very significant lower power consumption. This is almost 1% of the power consumption of a GPU, and on the other hand, much higher compute efficiency and lower latency. So really amazing results that they got when they're comparing to an Nvidia GPU, and this is really very encouraging. So we're seeing these very good results right now. We are focused on moving forward in this research work and getting to the point where we have really a strong solution for the market. And it's really very exciting. So if we go to the next slide, Weebit today has the embedded non-volatile memory used as a standard memory. We're now already working with people on using this embedded memory as a near-memory compute type of solution. And this is something that's basically available now. It's just using the memory in a different way, but it still is a non-volatile memory that's used for AI, and that's up and running already. And there's a lot of work now in commercializing in-memory compute, which I think will have an amazing impact in the market once we get it to work. So going to the next slide, we can see if you look at the three technologies that are basically being used today, ReRAM is the best technology for this combination of activities, both for storage as an embedded memory and including near-memory compute, but also as an analog compute engine which can store multiple levels in a chip. So it can work really as storing that coefficient, and you have the scalability that it can go down to the small geometries. And we're very excited about all of this work. We have been putting a big emphasis on the in-memory compute solution, a very strong team that's driving it forward, and you can see the initial results. So when we, you know, if we look at, I think it's my last slide of what my summary is, AI is really creating now new demands. Inference is becoming the main activity in AI, requiring really low power and low latency and very high density. ReRAM is great at addressing these needs, and we're seeing that today it expands ReRAM's role, the AI abilities over time. So now we're talking about embedded memory, we're talking about near-memory compute going to in-memory compute, and eventually to the holy grail of neuromorphic computing where the memory just mimics the human brain. So that's basically where Weebit is. A lot of interesting things going on, and very exciting times in the AI world.
So these agreements are very big, very elaborate. They take a long time to negotiate. You know, literally sometimes it's years that we're talking to these people, and they test and they verify. This is a major decision for a fab when a fab brings ReRAM in, when it changes. Flash has been around for decades. You know, replacing it, the thing that you're so used to and so confident about, with something new. You know, you can imagine the amount of testing and checks that they want to do. When we sign that agreement, we go into the technology transfer, and we need to basically teach the fab how to manufacture our technology. That takes time, and you know, they learn it, and then they need to manufacture a first sample, and then, I mean, manufacturing takes time, and they test the samples, and eventually they go into qualification, which again is several months, quite a few months' worth of work. So the overall until the fab is ready for manufacturing is roughly, you know, a two-year period. Now in parallel, we need to bring on board the product companies. That's where, you know, people like Onsemi, it's really good because they're built into that deal already. But customers of DB High-Tech, for example, we need to bring them on board separately. And those guys, again, we need to negotiate. Those negotiations are faster and easier, but still they don't take a day or a week. They take the time the people want to test and verify that the ReRAM will fit their specific application and their specific needs. You know, then they go into designing their product with the ReRAM, and that, depending on the product and its complexity, it can take from half a year to a year and a half or whatever. And then they need to manufacture samples, etc., etc. You know, when you're talking about semiconductors, unfortunately, everything takes time. The good news is that once a fab decides to work with us, this is literally a decision for decades, and that's what people need to understand. A fab will not go through this huge investment in time and money and people. You know, the teams that are working with us at Onsemi and TI, you know, we're talking about dozens of engineers full-time working with us. This is a huge investment, but once they engage, this is literally for decades. So it takes time, and that's maybe why they also hesitate and they take time until they sign.
But we are engaged with basically the vast majority of the manufacturers. I think the big manufacturers now, they do have a challenge, and you guys have been talking all day today about AI and stuff. The demand now from the fabs is crazy. Fabs are at 150% capacity. I mean, there's no way they can address the existing customer needs. So for us, that's kind of a challenge because they're so focused on, we just need to address the current customer needs. We don't have time now to really dive deep into this new technology and things. So things are taking longer. Obviously, I believed and I think everyone was expecting to see us sign another agreement already by now. It is taking longer, but I can assure you not a single one of these fabs has told us, oh, we're not interested in ReRAM. It's basically, yeah, we're just, come back in half a year. You know, we're just so swamped with work, we just don't have the attention span right now, and that's what's happening. So these deals, they are progressing just slower, and we will be signing them. And I think now as we progress, people are gaining more and more confidence because they're seeing now that in DB High-Tech things are moving forward. They're seeing what's happening at Onsemi. By the way, Onsemi and TI both expanded their agreements with us to additional business units and additional applications, which basically shows you that they dove deep into the technology, that saw that it's working, and now this is like new customers. Having additional business units in these companies is basically like new customers.
