Hi, this is Mick Elliott from Electronic Specifier. I'm at the Weebit Nano stand at Embedded World, and with me is Coby Hanoch, who is the CEO of Weebit Nano. First of all, Coby, can I ask you: why do we need a new kind of non-volatile memory? What's wrong with flash?
Well, I think it's already been over 20 years that people realized flash is starting to hit its limits, and ReRAM has significant advantages over flash. So, starting off with the fact that it's orders of magnitude faster, orders of magnitude lower power consumption, higher endurance, but I think maybe the two most important things are that it's cheaper to manufacture than flash. That's kind of a little important. And you can go down to small geometries with ReRAM. Flash basically gets stuck at 28 nanometers and you can't really have it in the advanced chips. So, that's a major advantage as well.
So, just taking up that point, do you work with the foundries on process, talk to them? Because I know you license the technology, but how do you work with the companies that are actually going to make them?
So, Weebit actually has two kinds of customers. We work with the fabs, the foundries, the IDMs, and basically transfer the technology to them so that they know how to manufacture it. And we work with the fabless companies, who basically embed our technology into their system-on-a-chip, and then they go to that fab and manufacture. So, we do need to teach the fab how to manufacture the technology. There is that technology transfer, but at the end of the day, as I said, it is cheaper to manufacture than flash.
Okay. I mean, I think there's two approaches to resistive random access memory, and you choose OxRAM. What are the benefits of that?
So, there have been all kinds of approaches to ReRAM. I believe that today the ones that are actually managing to work and getting to production are the OxRAM ones. It's just a simpler technology that, without diving deep into the technical aspects of it, actually ended up being the one that works the best.
Okay. So, you've made some announcements at the show about AI in-memory compute. How does ReRAM help with computing memory, the neuromorphic approach?
Yeah. So, ReRAM is actually an ideal solution for AI. It can go down to the small geometries, and AI is mostly below 28 nanometers, so you need something there. And ReRAM is an analog-type memory that lends itself to the needs of AI.
So, I mean, from the simple thing that people are doing right now: if you have something at system-on-a-chip that is smaller than 28 nanometers and you're doing inference, for example, you need a separate chip, a flash chip, to hold the coefficients. And then when you turn the system on, you need to copy everything from that flash into the AI chip, and only then can you start running. So, two chips are obviously more expensive than one. You know, the boot time takes a long time. And while you're copying, you have a security issue. So, there are a lot of issues there. So, just replacing the internal SRAM, or part of it, with ReRAM, you have a one-chip solution. Now it's a non-volatile solution. It's instant-on. You remove that security risk. The chip itself becomes non-volatile. Now you can turn it off if the sensors aren't sensing anything. So, a lot of advantages, and that's even before we start talking about in-memory compute, where it's really an ideal solution. Weebit has been researching it for more than 8 years already with different research institutes. So, AI is actually a very, very exciting domain for us, and we expect to be significant enablers in that space.
Okay. So, can you just give us an idea of what you're demonstrating at the show here?
So, we actually have an AI demo, a gesture recognition demo right over there. It's a demo that was prepared with an actual product company, with EEMITS. They used to use MRAM in the past. Like many other companies, they've been switching to ReRAM and using ReRAM for it. And so, this is a very nice AI demo that you can see here.
