#503 – Fabless Chip Design with Mohamed Kassem

Mohamed Kassem: This is The Amp Hour Podcast. Released August 2nd, 2020. Episode 503. Fabulous Chip Design with Mohamed Kassam.

Chris Gammell: Welcome to the Amp Hour. I'm Chris Gammell of Contextual Electronics.

Mohamed Kassem: I'm Mohamed Kassam, co-founder and CTO of eFabulous.

Chris Gammell: Welcome. How are you doing?

Mohamed Kassem: I'm very good, thank you. How are you?

Chris Gammell: I'm great. This is so fortuitous timing. We had planned to meet, but then some of our scheduling stuff, some of my schedule stuff didn't work out with another guest, and you were able to join me here today. And I'm really glad you are, because the timing couldn't be better, given the fact of the coincidence of Tim being on the show last guest episode, and obviously the Google PDK, OpenPDK stuff with Skywater and eFabulous doing a lot of the... You guys are doing the handling of the 40 designs as well, right?

Mohamed Kassem: Yeah, that's actually a part of the picture. So eFabulous started by actually releasing the OpenLane. It's an environment that people can use to build these chips for the slots that are going to be in the shuttle, both sides, front end and back end. And we're actually in the process of publishing how this is going to work. And people are asking how and where, where do I go to submit my shuttle? And they need more information. So we're working very hard to do that quickly. Great. Because the community is actually very engaged. So it's a positive thing to get here.

Chris Gammell: Yeah, that's great. That's great. So let's take a step back. For people who didn't maybe listen to the last guest episode, could you give us a quick rundown on what some of the news is and what's been happening in this world? And then maybe then after that, we can go talk about how we got to this point.

Mohamed Kassem: In the worlds of semiconductor in general?

Chris Gammell: More about the OpenPDK and the Skywater stuff and Google and all that.

Mohamed Kassem: So just in the general, in the semiconductor industry, it had historically closed in terms of the sharing intellectual property and knowledge. They're contained in pockets within companies or subgroups or universities. And in order for someone to be able to, even a small company, to be able to make their own design or just choice of design if that's a choice to them, they need access to the foundry. They need access to IP, like functional blocks that they can put on the chip, interfaces, et cetera. And they need software, EDA software. They need maybe complementary knowledge that they don't have. So, and testing capabilities. So if you put all that together in a one big end gate, that's what I call it, it's zero in the end. If any one of these don't exist, you can't make a chip. So it is, as we go forward, if you notice in the news in general, the companies, semiconductor companies have been buying each other and the number of semiconductor companies is going down. And that means there's less flexibility for the agile people who are very knowledgeable, but willing to take risks.

Chris Gammell: Yeah. So like if you want, yeah, if you wanted to like differentiate yourself in the market, it would be really tough. I think as, you know, we talked about this last week about analog buying Maxim. And then you talk about like, how do you, how do you differentiate yourself from someone who, another small company that's using the exact same chipset. So it's like you have software, but if you want to differentiate yourself at all on the hardware side, it becomes very, very difficult. It seems like.

Mohamed Kassem: Yeah. And in order to be able to, you know, just historically, just the idea of having a custom chip, it rings a bell right away of half a million to a million dollars of a start. So people, people are just like, I don't, I'm not going to think about it because that's the way it is. And not knowing that originally when I used to work, I used to work in the, with Texas instruments and the, and wireless engineering group and the, the volumes of the product are, you aim for a hundred million units. So, which is the investment would be no matter what it's going to be too small. So you do a custom chip for that. Now it became this way and the end gate became an obstacle for small general, small innovation and for small ideas to come to the, or new ideas to come to the, you can try to prototype versus market. There were open source softwares and EDA, they existed in, around the ecosystem, but they're different units or separate engines. They're not a flow. So you would have every, every, as a simulator, the synthesizer, the layout checker, every one of them is doing something and it's working on their own world. So there was a need for something, for something to happen to aggregate or to glue all that together into something that other people can use in a, in a use of, in a useful way to actually get some silicon out of it. Just as a background, and I'm leading to the Google part here is as a background. So when eFablet started, we, the, the approach is very simple. We wanted to have in the browser, give anyone the ability to design, yeah, any, anything they want on a, on a custom chip and provide it, put it in a marketplace that a customer or a collaborator can work with them on. And when you think about just what I just said, it implies two things. It implies that the end gate, all of it would exist in the cloud, in a, in a portal where someone can, I can, I can communicate with someone for knowledge. I can use the EDA software. I can access the foundry. All of that is in one place, one cockpit. That, if you think about how to do that with a commercial or a traditional ecosystem, you're going to be very quickly in the impossible space, foundry, IP and EDA.

Chris Gammell: Yeah. Right. I'd say even just the EDA, it's like, yeah, I guess we'll just, our startup is we basically bought a hundred seats of mentor graphics and we just put it accessible online. And yeah, I mean, that's how you do a startup, right? You light money on fire. Yeah.

Mohamed Kassem: Yeah. So luckily I, you know, I had some understanding of some of the goals of the EDA companies back then when that was around 2012. And I, I approached them to entice them to, to have a change in the business model justifiably. They, which I, it's very well understood. It's that their economic model is not suitable for that. So I immediately started doing, looking or on a mission for open source engines and actually did something that's successful. So yeah, I literally traveled the world and I put them, you know, and get the right people and, uh, it was, had the honor of getting to the privilege of knowing them in more details and visiting them in our homes. Anyway, we got all that and we started putting this first stack. It was an open source environment and it didn't have an open source process.

Chris Gammell: Yeah. So yeah, and it sounds like, I mean, so you knew all this, you knew all the steps anyways, because of your, your past experience. And I'd love to hear about that at some point, but you basically had to build, like you're saying, you had to build all these more open source or really more low costs. It seems like is the, is the real key to make it accessible. And so you had to build up each tool and each layer, you know, not build it yourself, but build it up as a ecosystem. And then once you did that and the present day, it's like, there's enough tooling out there to, to actually make it accessible. What about the open PDK? The, the last layer actually enables someone to go and make their own chips now. So like, why was that the gating moment or, or, or was it?

Mohamed Kassem: Well, it was because, uh, if you think about it, the PDK is an, if you think about it as an intellectual property of the foundry. So, uh, if the foundry is, uh, protecting it and requires you to protect it, then you have to. So the middle ground was, we put the open source solution on the cloud and we put the foundry PDK on the cloud as well, but obfuscated. So, uh, so people can use it from the portal, but they cannot download it because it don't download implies full exposure. So we worked out with the two foundries, at least to have that agreement with them in such a way that you can go to the website right now, log in, you can use 130 nanometer on global foundries or 180 nanometer next fab. And very soon it's going to be 130 nanometer from Skywater. The difference is now when the PDK is open, it's exactly what, what you saw here. I can go to good hub, get the PDK built and I can go to get open lane built. And then on my computer, I have the ability to run both and do a lot of iterations. By the way, on both sides, the software and the hardware. So as you, as the minute you publish these things out, people started testing and trying their own things that you wanted to try. And then someone would add a feature and say, okay, it would be good to add this to the software to open lane. Okay. And, or, or find an issue, missing information in the PDK. So there was a lot of open collaboration. So when you think about the two worlds, I have to collaborate within the building and invite people to the building carefully versus I'm collaborating in a distributed world where I am sitting on Slack talking to people that are in different time zones. I have a goal is to get this environment to produce chips, to put them on the shuttle in November. Yeah.

Chris Gammell: So it seems like it's, it's reducing friction at every step, including at the developer step. The thing that I guess I wanted to call out though, is so you'd mentioned a lot of this, you wanted it to be in the browser. The fact that it is downloadable though, how much, what is the split of like, is it possible to do everything in the browser? Or is there a realistic, if you really want to customize things, it has to be on your local computer.

Mohamed Kassem: So the browser versus not is, the browser is nothing more than a delivery window. At the end of the day, I call it a remote display. So it's whether it's in the browser or some other X client that you use locally. So the challenge is typically with that is internet latency or data center to, you know, so the latency to the right place in the U S in general, you wouldn't see that problem. Uh, you start seeing it, for example, in India, there's our, so we, we end up deploying or looking at other data centers for the cloud provider that we're using. That said, because of that latency, you put restrictions on the size of the design that you're going to look at because it's going to refresh in a long time or, so these are things that are challenges. However, they haven't stopped people from continuing. We have people, you know, a lot of people from India, a lot of people from, uh, the Middle East, Europe, and of course here in the U S, but they don't have that latency issue. They don't see it. So it's actually very snappy when you try it, uh, from a good internet, low internet latency. It's not about the end. It's more, it's at some point at the latency, it makes a difference. Now that is not the issue for the open source side.

Chris Gammell: Yeah. It seems like a nice to have versus like a must have kind of thing.

Mohamed Kassem: Yeah. The only reason we had it in the cloud, it's simple is the foundry and other IP providers, other blocks were trusted. They put them available in an obfuscated way. So the community can, on the other side, use them without seeing what's in them. So that, yeah, if fabulous was like black box kind of idea, like inputs and outputs, transfer

Chris Gammell: functions, all that kind of thing.

Mohamed Kassem: Correct. And if this was the kind of a straddling, the two medium, uh, one in one leg in the, and it's still, we do. So now we have companies that don't want to open their IP, but however, the open source community can upload their design and hook up to that IP. And so we're straddled that medium. Now, when you download it, you can do everything off the platform. And then when you're done with the design, we have a process that will absorb it from your repository and check it against a set of checks so that it becomes manufactured.

Chris Gammell: Yeah. That's great. That's great. Before we dive into the flow, because it does sound like it'd be really interesting to kind of walk through the flow here. I mean, we talked about that a little bit with Tim, but I think that it would be a better, a better to, to kind of do each step here with you. How did we get to this point? I mean, so you mentioned 2012 is when you're starting this, but why this? Like what, what made you decide to go in this direction in the first place as someone who's had a lot of experience in the industry?

Mohamed Kassem: So I always, I invest this as personal personnel. I always enjoyed working on new things. And even when in my previous life, before I started a company, I was working in the, in the smartphone chip development area, which is back when from 2000 year, 2000, this was a height of the smartphone. And it was very, very good opportunity to learn. And it's like a bootcamp. So I learned a couple of things is that, that you, you can design, you connect, you may not depend on the tools so much and you still can design because we had these issues inside the company and we got around them. So I learned a lot of the, how do you do it with a low confidence or if you have a low confidence in the tools, for example, or the models or. So one of the things that I looked at and actually saw it very obvious in front of me is that inside the companies that develop analog chips, only like smaller, like the, you can actually have two engineers inside the company responsible for a complete product. And the rest of the resources are coming from the company, whether it's the infrastructure tools, access to the fabs, et cetera, testing, all that is infrastructure. So it hit me. That's just the thought. And I'm not sure. I'm sure it's not necessarily a unique thought may have come to many other people. And, and by the way, this is actually a spectrum. The chip complexity is a spectrum. So there's chips that we can actually, one person is developing and there, and there's a chip that has, you know, 150 people on it. Okay. So, so I looked at the other spectrum, the right side of the spectrum. And then I, I found, is there a possibility to produce something or replicate a model that looks like the Apple app store in a way where people can design at no cost or, and get a development kit of some sort and then deliver in a marketplace. So that's the general concept. When I thought about it, it started to evolve very quickly. And, um, I decided to take a risk on it and just, I wanted to actually do it cleanly. So I, I, uh, I, I didn't even have the Ford formulation. So, and I left TI to try to solve the problems one on the end gate one by one, including the foundry.

Chris Gammell: Yeah. Yeah. That's a lot of steps to get there for sure. On the, uh, on the app store side of things, I mean, did, had you been talking to engineers who wanted this sort of thing or had you, had you seen it or is it more of a kind of a feeling that this technology's time had come and you're just kind of ready? You feel like if you build it, they will come.

Mohamed Kassem: Well, it's not exactly. It was an intersection from a few points. So I was looking at places in the software world, like a company called top quarter. It's a pioneer in community-based software development where you can throw up a problem. It could be a problem, data set problem, data analysis problem, all the way to building a small app. And then the community will actually come in and, and contribute and develop that for companies, for like IBM, for example. So there are precedents in the, in that model where the community can, uh, in the software world. And then I noticed also the, the, you know, I had, when you say the feeling that data points around the open source world where people are accessing the software, but it's always stopping short of a complete design that can go to manufacturing.

Chris Gammell: Got it. Yeah.

Mohamed Kassem: And I thought this is, I basically the, the piece that I thought that it's going to be, I'll build a, you'll, you'll get there is when we put out the solution, people will come because it's a valuable solution and it's not offered before. And yeah, that, and that was actually the, um, it was a, but Peter with a gut feeling and a gut feeling is developed usually from your experience around talking to people or seeing how it works. But there was no definitive, like, Hey.

Chris Gammell: Yeah. No one walked up to you and said, Hey, Mohamed, could you, could you please design this, this, this market ecosystem for me? That would be, I really need it. If you could just get it, get to me, get it to me whenever you get a chance, that'd be great.

Mohamed Kassem: And I have to say we, the learning, I mean, you get to learn by fire and everywhere and you try a lot of things and you take feedback from people. And so I call the, as we evolved since then, it's a lot of it is by learning and working with a lot of people.

Chris Gammell: Yeah. I'm sure there's been a ton of learning. I mean, um, so one thing that's interesting to me, and I think I might ask Tim this as well, but like, so like from my personal experience, I, I didn't see, I don't see myself making a chip anytime soon. Right. But that's also because I'm small potatoes. I'm not working on, you know, high volume products. What is your, what is your take on when it makes sense to go and push some custom silicon? Like, what is the volumes that it, it takes to do this? Because maybe it's just that I'm not even near those volumes. And, uh, but some people are, are often working in that world.

Mohamed Kassem: So the volume question mark came from an interesting, from just a simple economic formula, which is if you're going to spend the money and resources to build a chip with engineering, you need to sell enough to justify that money. That's, that's simple. Yeah. And if you're still, if you're selling one unit, let's say, imagine you're selling one unit, then the entire engineering project budget would have to be converted into the baseline for the chip for one unit. So imagine that. Right.

Chris Gammell: That's the first chip costs 10 million. The second chip costs 10 million and one, one dollar or whatever.

Mohamed Kassem: So now that said, the world is now, when you talk about chip and ASICs very quickly, they jump into the advanced technologies, 40 nanometer, 28, 22, you know, it's 14, 12. And when you hear these numbers, the upfront investment, yes, tens of millions. And in this case, it doesn't make sense. You have to have the volume. You have to. And the smartphone, when we worked on them, you have to, because the smartphones were exploded, have more than half the planet have smartphones, have phones. We wanted to convert the formula into change that, into remove the volume question mark. And how is that? If you reduce the barrier to, for someone to get the prototype, there are a couple of things that can actually prove something and they can get investment maybe. The number one question is people will ask, do I need an ASIC? Well, the trade-offs between an ASIC versus two chip solution or some other multi-chip solution is something that knowledge is residing typically on the ASIC side. And that means that a system developer, they wouldn't necessarily contemplate or figure out, you know, just to quickly say, okay, I can do this on a chip. All right. Maybe they would. But immediately the economic part kicks in and says, well, it's going to be too much.

Chris Gammell: Right. Yeah. So it's like having a tool in your toolbox, but that tool, the nail costs a million dollars. And you're like, well, I'm not using that nail because I don't think it makes sense.

Mohamed Kassem: Exactly. So if you change the economics and say, like, for example, in the high volume world, you have to make the device area very small so that you can maximize the number of devices per wafer. And in this world, I say, well, is area so important if it's a low volume? And if it's not, then I can actually relax some other component and change the formula here, at least change the NRE, the amount of money being spent up front. So if you reduce that amount of money by providing tools and by providing the access to the foundry, then that budget that I've said it would be divided on volume would be divided on a small amount, a smaller amount, much smaller amount, like you're talking about 10X.

Chris Gammell: Yeah, that's great. And so, okay. So then let's play with that though. So you've now gotten the cost down, you've got the chipset that's available, it's accessible, it's easy to use, all things that it seems like we're driving towards. But then the reason to do it is what, security or power? Or like, what are the reasons to do it after that then? Or uniqueness?

Mohamed Kassem: So there are multiple reasons. You mentioned at least two of them. So in cases where you'll be able to lose lower voltage because of the chip, or you integrate two chips on the same guy, you change how the IOs and all the power requirements work. So the power is something, but it's not necessarily a given. So it always has to be analyzed because there are chips that are consuming more than systems. And the other side is security. Security here is kind of interesting because I call it protecting your own IP. So if you have a specific algorithm, we're working with a company that has their own specific algorithm to manipulate data coming from the sensors. And this is something that they use it for their product. So they have a full product and they do the chip. They buy the chip. So they wanted to replace it for supply chain reliability because they only had the sole source and they didn't want to be dependent on it. So they wanted to initiate another design to create a mitigation of the supply chain.

Chris Gammell: If anyone's going to mess up my supply chain, it's going to be me, darn it.

Mohamed Kassem: Well, yeah. And basically, they're working with another company and the company is great and everything. But he basically said, if that disappears, and he looked at the investment, he came to us and looked at how much money is there. And then it made it easy. Coming back to the question was, so the power, the security, or protecting IP form factor and just pure volume like space is actually miniaturization is when you need that. When the product is at the level, like we were at the back in the smartphone days, you just say you need the chip to be no higher than that. The height, not just the area. Right.

Chris Gammell: Yeah, you can't make some four millimeter tall sip or something like that.

Mohamed Kassem: Yeah, exactly. And so another example is just an earpiece. So now how much do you want to put in if you couldn't actually put this electronics inside a chip that gets in there? So the miniaturization is an important aspect as well.

Chris Gammell: Okay.

Mohamed Kassem: The last thing actually is obsolescence mitigation, which is some people would have a complete board that is ancient. The parts don't exist. And you could have the option of designing. Somebody would design it again. It's either by replacing it with some multi-chip solution again. Or if you know that the trade-off with the ASIC will make it easy, then you can make that decision. And that's actually one of the things we're doing, trying to eFablish, is to make that judgment easier. So we give cost numbers and information as you are choosing what to be on the chip, what to exist on the chip.

Chris Gammell: Yeah. I can imagine even with the upfront cost, there still is going to be upfront costs, which we'll get to. But even that, you balance that against the last time buy of buying 10 years worth of parts, well, it might be worth it to go and spend the NRE upfront and make something that you can make for as long. And then you have the design files effectively. So you can make it as long as you find a fad that will push out the chip and validate it for you.

Mohamed Kassem: Yeah. And I have to say that if you really want to speak of it in realistic numbers, if you're designing your own chip, like you're actually doing your, bringing your own time and effort to make your own chip, the number goes down below. I mean, it could be well below $10,000. Okay. If you want to, in some process technologies. Okay. So it's a credit card question mark. Okay. And people will hear that and say, well, what kind of chip is it? Does it have 30? Does it have the USB? Does it have, well, no, no, it doesn't. Okay.

Chris Gammell: No, it's got, it's got what you need, but not much else. Exactly.

Mohamed Kassem: And so you can go that. An example with the Google shuttles, what's, what's happening now is that people have the tools, but that number that I was talking about, the 10,000 is gone. So because Google said that we're going to sponsor the shuttles so people can use it, prototype. So that, that is one component. The other component, which is interesting, that's kind of hidden is that if you're, if you have an idea and you believe it's going to be super useful and a lot of people are going to use it, the, if you go pitch, you don't have the money. Okay. So you want to get investors, for example, you go pitch it. And the, the, the, the slide where plus your experience and maybe the team will be the only indicators, whether that is executable or not, can be done or not. Of course, it can be done, meaning that you deliver on the chip, assuming that there's something that's hot in the market that people need. Yeah. Right. How many people get filtered at that stage? You know, I have an idea and, and I will do make it. And then, no, you won't make it. We believe, we know, we know what you want to make. I mean, that's how many people have told me that we won't make it. And we want to basically be able to run open source tools or any of that. So by having an upfront cost, for example, you come in on our platform, on the platform or actually open source on the open source world, you can develop all the way to finished chip. That actually is a lower risk stage. So if I want to go to the investor, I can actually say that I have this. I've developed it. I have a pre-silicon. I've already done it because it was, everything was open source. So it's, so it's, I have, I cut a lot of the risk for you. Yeah.

Chris Gammell: Right. It's like, it's like having the access to like an AWS server now versus, versus like how it used to be. Would have to build up a small data center and prove out your idea and have the stack and all. So you're saying that now you can just go and rent an AWS server effectively and, and do that for a ship. That kind of idea.

Mohamed Kassem: Yeah. And then in that case, that $10,000 I was talking about, even if it's 20, then it could be the customer or the investors. It's like, oh, I'm willing to do that. So I can see whether this is going to be, uh, you know, realistic or not. But I'm working from a real design, not from a theoretical design. So that's an important to move the money from the idea to the prototype.

Chris Gammell: Yeah, that's a good idea. And yeah, and again, you don't have to invest in even the, the tool chain and, you know, beg, buy and borrow the. Your own time. Yeah. Yeah. It's just, yeah, it's your time. It's your, it's, it's going back to a garage startup where you're actually, you don't have to, you don't have to fund the, the, the mentor license or the, to, to, to be working in that garage. Yeah.

Mohamed Kassem: Yeah. And this is actually a, we have good examples. Like we have the Raven, uh, microcontroller that we developed, uh, earlier. That is a, is, is a great example for, uh, collaboration. So we, the, the, the CPU developer is somebody from, you know, clear, it's a wolf from, uh, Austria and they had the CPU. They don't have access to the foundry or the IC tools. Uh, uh, we have access to the foundry with, uh, analog IP. We have system integrator. So it was built in a collaborative way, uh, over, uh, the, over the cloud. And when, and we ran it and we made it to Silicon and we published it and people now have their own work, the open source work. First time on a real silk versus an FPGA. Now, whether or not that's a cool or why you would do it? Well, it is actually, it said this, it actually opened the door to encourage people to say it's actually possible. I mean, when we did, the team did, um, and the, a microcontroller based on a risk five CPU with a hundred percent open source tool set. It invited, oh, there, there is something there. Right. How is the performance comparable? Maybe not. Okay. But is it enough? So that's, that microcontroller runs on how to megahertz. Okay. And is it enough for some applications? Yes. So, and this is where, you know, that was the pre the Google discussion about, uh, getting the opens, working with the foundry on the open source process.

Chris Gammell: Yeah. So let's, I mean, that's a, that's a great, uh, great transition, I guess, into some of the stuff that's on your site. And then maybe we could talk about the flow as people are using it and deciding to go towards, uh, you know, making it a chip. So, so you mentioned the Raven microcontroller, the Claire designed. What is it that people are going to do each step to get towards making, making their own chip? So if they're, they're in the garage, they're like, all right, I'm ready to prove out this $10,000 design to this investor. What are all the things they need to do to get, to get to the actual silicon in their hand?

Mohamed Kassem: So the fact that, you know, if you think about it, the, if you think about the end gate, when you work on something like on a fabulous platform, everything in the end gate is available. Now, the most important part is knowing, uh, is the knowledge. So if you're doing it yourself and you have the knowledge or we have, you have someone complements you on that knowledge, just go to the website. And not only will you find the tools to start from, you'll find functional proven silicon IP. Uh, again, you asked me what they are, are they USB inserties? I'm going to say no. Uh, but, but they're actually proven coming from the foundry. And not only that Raven is as a chip integrated and tested together is available for you to clone it and make modifications to it. Instead of building it from scratch.

Chris Gammell: Yeah. Yeah. That's, that's one thing I had asked him about as well as like a starting point to, because that, that is so often, you know, like that's how I learned a lot of things is, you know, I start from a development board effectively. And if there's nothing like that in the silicon world, then what? Exactly. So Raven, so we have a board for it.

Mohamed Kassem: And that board in people who partnered with us have used it. And as a matter of fact, there was a customer that took it. And the second, a couple of days after we published the Raven, they were interested in it. So the board existed. So not only can you touch it, you can, you can experience, experience it. You can understand the peripherals and know what it is, but you can actually clone it with even verification benches and things that help you attach other peripherals and things you put in. So that reduces the time first, the effort. So if I am, if I am an analog designer and I know a certain focus and I want to develop a state of the art or some new architecture of an ADC, for example, I can use the rest of the chip just by unplugging that and putting that, putting mine. So, and there is no cost for the chip itself as an open source design. It's up to public out there. Yep.

Chris Gammell: That's great. Okay. So they, yeah, they start, they start with the Raven and then basically they're downloading the, the Verilog effectively and then stepping through with open lane. And like, what are those steps?

Mohamed Kassem: Specifically Raven, it was done pre-open lane world. It was done a couple of years ago and it was done with Keyflow. It was managed by, it's a, it's a, one of the star flows that are coordinated and maintained by Tim Edwards of opencircuitdesign.com. And so that's what we used. And then Raven had an interesting missing part is that the IP inside it, not all of it are open source. So although you were going to find RTL. So actually when you clone it, you'll be able to get the RTL of the top level and RTL models for anything inside the chip. So you can simulate it. You can simulate it. You can add your own. So you can simulate it. You can add your own, as long as you use the Lego blocks that exist proven that are on the platform that are provided, for example, by XPAP. Not an Raven. So there's a constraint in Raven because the process and the IP were not open source. So in Raven, you would, when you clone it, you're not going to download it. You can actually just clone it on the platform, put it in your own workspace and do whatever you want with it. And people, some people will consider that a constraint. Some people will say, I didn't even have that before. So another example I'll tell you, ARM, today we have an agreement with ARM. You can actually get to our platform, use the one of their cores, like the M0, without talking to ARM. Just right there, within minutes, you can actually be able to use it and put it in a system and simulate it and compile it, et cetera, without actually talking to me or ARM or anybody.

Chris Gammell: So you're saying that in order to use this kind of more closed off, if there's something that's not completely open, as long as it stays on the site, because it's behind some levels of obfuscation, it's no problem. Like having, you don't have the RTL of the ARM0, M0, you don't have the RTL. We do have it in our, you know, in the back end. Sure, yeah, yeah.

Mohamed Kassem: I mean, sorry, you, I was saying you as the user. And that type of story and setup made it comfortable for people to say, I can give that, put it here so that other people can use it, because eventually it's going to sell more if people use it more. But the risk is less because they're not downloading the actual IP. That's if the supplier or provider of the IP is actually closing it. Now, in the new world, so, well, actually, before we go to the new world, you clone it in your system, you simulate it, you verify it, everything is online. And then you request to be put it on shuttle. And then in this space, there's a payment there. And then you, you get back parts. Yeah, that's great. So you can go clone Raven, modify it, add something to it. We did the same around ARM M0. So it's available there. Now in our new world with the OpenPDK, we have a family called Strive microcontrollers. And they are completely open. They will be on GitHub. That's awesome. And they're built using OpenLane.

Speaker ?: Okay.

Chris Gammell: And so Strive is like a core IP block. It's like a RISC-V processor with everything around it. Is that the idea?

Mohamed Kassem: It's actually, it's the first one is going to be the core RISC-V processor, the Pico RV32, because that's what we used in Raven. So just, we wanted to minimize new factor. So we jumped into that. That's going to have a logic RAM and an interface, an SBI interface and a clocking system. There's no analog whatsoever on it. And the logic RAM, because back then when we did it, there was no, we wanted, we didn't have the compiler ready, the RAM compiler. So we have another one, Strive 2, that exactly the same content, except that the RAM is not logic. It's a compiled RAM. Actually, you'll see it in the chip. Like a block versus just the logic integration.

Chris Gammell: And then we're adding other things. Can you explain that a little bit more, actually? I'm a little confused about what you mean there.

Mohamed Kassem: So the RAM, you can either implement it literally, like with a bunch of flip-flops on your own, lower performance. Oh, you're saying because it's SRAM versus DRAM? Well, yeah. But you can use the elements of the library, the flip-flops themselves. You can use them to construct what looks like a RAM from the outside. Yep. Yep. And that means it's not regular. It's performance is lower, but it's a RAM. It's actually a legitimate RAM.

Chris Gammell: Right. It's like it will store and randomly access. It will store and you can have access and all of that. Yeah.

Mohamed Kassem: Now, that is, if I show you the two chips, you're going to find that this one is bigger because you're implemented this way. The other approach is that you specify the RAM size and parameters, what you want. And then you get a macro, you get a block or a rectangle. It says this is eight kilobytes or one kilobytes of SRAM. And you drop it outside the CPU, outside the chip, outside the main parts of the chip. And if you look at Raven right now, you'll see it here. If you go to the marketplace and look at Raven, you'll find it's on the bottom right corner. A separate block.

Chris Gammell: Okay. Okay. So it's something that you, yeah. Yeah. So it sounds kind of, it kind of reminds me of like, in like an FPGA, you can have, you can use the memory element of a LUT or not even a LUT, but a logic cell, I guess, inside of an FPGA, you know, logic cell unit, whatever there is. There's just a little bit of memory. There's a LUT in there. You can implement it like that. Or then there's also block RAM that you can access. And it's just like each vendor kind of implements it different, but there's like this set amount of stuff inside the FPGA that you go and instantiate and say, no, this is, this is a special type of memory here.

Mohamed Kassem: Yeah. Yes, exactly. And the Strive, we actually have like now six versions of it, but because, you know, because we're engineers and we want to take, be logical and systematic. So the chips that the Strive family, which by the way, I'll be talking about it in August 25th on the Fosse Foundation dial-up program, I'll be presenting the Strive family. Yeah.

Mohamed Kassem: So it has, every chip has a different new thing that we did with open source. So the first one is done with open name. The second one is done with open name and open RAM. We used open RAM as the compiler for the RAM. So that's a new, new, new element. A third one, we didn't use the library, the same library. We used the library developed by Oklahoma State University, completely different library. It's the same function.

Chris Gammell: So just to prove that it can work on, it's not just this one dialed in thing, it's the process that can handle a lot of different types of inputs.

Mohamed Kassem: Yeah. And the good thing is that all of them are Strive. So they're expected to function the same.

Chris Gammell: And when you say Strive, you mean like that's just the brand of like the overall chip design?

Mohamed Kassem: The family, we call it Strive. So now the first one has the logic RAM, but it functions a certain way. The second one has a block RAM, but it also functions the same way. So I can actually have apples to apples comparison on the silicon when I put them together. Got it. From outside, from the outside world.

Chris Gammell: That's great. That's great. So to go back to the flow, so okay, so now let's just say they're cloning the newer Strive thing. They download everything that's on there. It's got the open stuff. They put it through Yosis and then OpenLane is to do the place and route. Is that right?

Mohamed Kassem: Sorry. So OpenLane is actually the combination of a few things. OpenLane is intended to take you from RTL all the way to the GDS. All the way. So it includes Yosis. It includes OpenSTA for timing analysis. It includes something called fault for DFT. And then it includes OpenRoad for floor planning placement. Oh, that's right.

Chris Gammell: Yeah. Sorry. I was confounding those two terms because they're... So OpenLane is kind of the overarching flow for everything. OpenRoad is the one that I think Tim mentioned as well. That was the one that the DARPA had worked on, right?

Mohamed Kassem: Yes. And the OpenRoad is just... We have to say it's a huge departure from the world of previous flows in the backend. Like you couldn't do beyond 100,000 gates to get them into GDS. We were able to get into a million with one shot. So... And this is actually... This flow is very capable of advanced nodes as well. So the UCSD team is working on a chip for 12 nanometer from Global Founders. Oh, wow. Okay. So it is actually a capable backend and we intend to use it everywhere we can. Yeah.

Chris Gammell: Can you explain what GDS is? Because we didn't really talk about that. I don't think we've talked about that on the show before.

Mohamed Kassem: So the... A GDS is a format that is efficient in storing the data that you'll present it to the factory. So the closest example of that is simply the 3D models. So when you go online and you have a 3D model, 3D model, just the STL, for example. STL is a representation of the 3D model that you're looking at for the printer to use to print. Yeah. This is exactly similar. So GDS is a graphic design system. And it has the data that represents the 3D aspects of the chip because every layer... The chip is a 3D-ish element. It has a... You build metals and you put it on top of it layer by layer by layer. So it's actually kind of... You can make a very, very tiny 3D printer. Right, right.

Chris Gammell: 300-step 3D printer with... Yeah. And literally, yeah.

Mohamed Kassem: And it has a combination of itching and additive and subtractive mechanisms. But it is essentially a 3D printed element that requires very high resolution to be able to print it. So that's the GDS. It's just a representation. And the data could get high, so very large in size. So there are multiple options and things to actually make it more efficient storage and reduce it.

Chris Gammell: Interesting. Yeah. So I guess when I think about it... So when I think about it in comparison to a PCB, is it like a Gerber where it's like a flat format for each layer? And then you say... But then you also need levels in between to say, well, actually, even though here's the pattern of the oxide layer that we need on level 30, then level 31 that you're doing as an additive layer, we need that to be at least 100 nanometers of oxide in order to get these certain properties.

Mohamed Kassem: So first of all, in the GDS, the standard practice is all of the... Imagine the Gerber layers, the different Gerber files. They're all combined in one file. Okay. That's the number one. Now, if you look at the PCB, for example, things like their layer thickness and their electric constants and things that are in between metals. This is for a given substrate or a manufacturer, it could be fixed. So when you go to the foundry for the chip making, the oxide thickness is pretty much set into a few standardized, some for the transistors and some for just filling what you call field oxide. But once you provide the GDS and I do layer one, it is kind of known for that process what to do to basically deposit the deposit oxide and the... Or not oxide. The deposit chemicals to get the oxide. And then you add another layer by deposition, like the metal. You just deposit it through the mask because it goes through the mask. So it gets added layer by layer. So the GDS is one bucket and then you take it, they split it into several masks. Every mask has one layer. And then they use it to do like a silk screen printing on the layer below.

Chris Gammell: Yeah, positive or negative as well, right? Do you need to take stuff away? Do you need to add stuff? That's like how you need to make sure that like you're not making a hole in something when in fact you actually wanted to fill everything around that one area or something like that.

Mohamed Kassem: Exactly. And when you said oxide thickness, this is kind of a, becomes a parameter that is controlled in the process. So they want it to be to a big degree constant. Nothing at this level is absolutely equal. It's always statistical process. But yeah, but the oxide thickness is known for this technology, given technology. Got it. Got it.

Chris Gammell: Well, speaking of, so a lot of the stuff that you show on your site is 180 nanometer, but then you also referenced earlier the global foundry is 130. And also, I think you said 80 nanometer. So, and then you said Skywater as well as 130. So what is the current access that eFabless has and then what just generally people might have in the future?

Mohamed Kassem: Right now, immediately, as it is today, anyone can access the eFab2 process technologies, 350 nanometer for high voltage and applications. It's a very good process for that purpose. And then 180 nanometer. Immediately you can go there. And not only will you access the process, you will find blocks like data converters, comparators, PLL, things that you can actually use along with your own design. Yeah. So that's, that is ex-fab site. Today, existing on the platform is global foundry is 130 nanometer, 130G specifically. And you can design with it. And it's actually the front end. You can do the front end with it. The back end is going to be based on OpenLin. And it will be released in a couple weeks. You can access the front end of the, of the global foundry today. And then within a couple weeks, you'll be able to actually use the entire process.

Chris Gammell: Huh. And so front end versus back end, you're saying is like, is that defining like the transistor level versus like the metallization layer?

Mohamed Kassem: Well, that's, that's actually, that's a one way of doing it for process people. It is back end of line, front end of line. You're right. Now, the, in the, in the, in the terms of the design, it's anything before synthesis. Once I start targeting the technology itself and mapping the design into specific technology, then becomes the back end starts. So anything RTL, very log, all of the system, very log, anything before that is front end. And then synthesis is the bridge. And now you have the design is represented in components that are specific to this process. So now you take that and place, placement, flow planning, placement, router, et cetera, all the way to the end.

Chris Gammell: Okay. Okay. And then what about, so you mentioned, I mean, the Skywater PDK as well. So like, can you walk us through real quick, like what it takes to get a new fab onto your platform?

Mohamed Kassem: So there are two, two to three things, but, but there's the business aspect of it. So that's basically making the case for the, for the, the foundry to believe that having this process on the platform that would potentially increase the number of designs or potentially increase the number of products that are being sold with as a founder. So that's a separate topic. Yeah.

Speaker ?: Yeah.

Mohamed Kassem: And unless you want to address it now, we can talk about it. And the second aspect is a technical part, which is the PDK is represented in a set of files that are compatible with the commercial EDA infrastructure or tools today. So generally a foundry would, when you ask them for the PDK, they don't give you something that works out of the box with an open source tool. Right. They'll give you, yeah. Yeah.

Chris Gammell: Right. Cause they're basically, I'm sure that the, you know, the, the large scale ADA companies are helping to develop this stuff because it's in their best interest to be completely integrated into a foundry business.

Mohamed Kassem: Correct. And, and actually, and this is a good model and it has its customers in its market. So it's a, it worked, it works in that, in that space. Yeah.

Chris Gammell: I mean, I, to, to make a, to make a, a, a crude comparison would be, you know, like I've talked to all these assembly houses. They're like, yeah, you know, we use all these different formats and we work really well with like Altium and, you know, Allegro and all these cadence things. It's like, well, thank God for spreadsheets and basically spreadsheets and Gerbers are the only reason that Kaikad can, you know, operate in that space is because there is that standardized format. And it sounds like now the chip industry is kind of starting to move in that way as well because now there's going to be open PDK and it's going to get a little easier to integrate new things.

Mohamed Kassem: And partially that existed in, like in the layout, GDS is a standard. I can read GDS with any tool, any of the commercial tools. Yeah. Spice knit list or timing models. Some of these things are like, however, there are not uncommon things that are enough to make, to create the stickiness, meaning that you're, you're going to use a certain flow, you stay with it. So now what we did to avoid, to be completely clean, we went to the foundry and said, give us the information, the process information, not even the PDK. Okay. And we will take it and produce the same thing that is produced for the commercial system to, to work with the competitive, with the open source tools. Yeah. And, and we take it from the foundry directly. They take the, the, the information, the process information. We don't even, cause we don't want to even, you know, be in our way to do any. Yeah.

Chris Gammell: Cause it's like interpretation almost of like data and how it all fits together. And PDK is like an abstraction of what's actually happening on the floor of the fab. Right.

Mohamed Kassem: Exactly. In different views. Some of them are physical views represent the layout and some of their simulation views represent the best basically functional and parametric simulation. Huh.

Chris Gammell: So then, so what is, so what is, what is, what is the work then that you have to do? So you take process information that's basically like what measurement of gate sizes and variants and statistical process control or what?

Mohamed Kassem: So to start with is for example, the, the, the design rules. So the design rules typically, they literally have a manual, a PDF manual, which is, this is kind of, this has changed already because in the release with the Google, but that PDF manual has every rule. And what is the metal spacing needs to be that much metal spacing, this metal overlap VSI. So all of that is written. Literally we take these numbers and stick them into a rule deck that is readable by the layout tool. It's just magic. And that's, so that's a, that's a start. Now, the good thing is you go, you're left, you go right, you have a baseline or a back end of line, as you said it earlier, transistors, and then you have metals. So the same, the physical effects all the way to call it 90 nanometer are very similar. So we, when we developed one deck, basically another 180 would look like, you know, a replacement of the parameters, but you don't have to change everything in the rules because the number of layers is similar. So that's, that's, that's one of the, one example, uh, in the, um, we literally get it from the source of information. Sometimes this information is, is encoded into GDS. Like I give you a GDS and say that GDS passes the DRC and the other one fails the DRC, the design rule check. So now I have an AB kind of, and, and you use them after that for testing, see if the tool will catch the, the, the violation. So we, the foundry gives us this information under, you know, because we, to, to enable us to do it right.

Chris Gammell: And then the, the validation that you're doing then, so you're saying that that's happening in a tool like magic or what are, what are some of the tools that you're using then to validate that you mentioned?

Mohamed Kassem: You're talking about validating the, the, the correctness of the PDK itself or the design?

Chris Gammell: That's yeah, that's right. So, I mean, you said you, you're kind of skipping the PDK and you're validating yourself. So you must be doing this in some kind of tools and like, how do you actually get to the point of like a source of truth, I guess?

Mohamed Kassem: Yeah. So the best approach that I call is that generally speaking to, to start from the foundry and get from the foundry, the, the way they sign off and test and stress test the rules in general. Basically, it's a, it becomes a bunch of layouts that have different variants of pass, fail, pass, fail. And literally when you run it through the checker, you'll see, you're expected to see 20 fails on these rules and 30 on this other combination of rules. If I don't see that, that means I'm not compatible with that. I mean, I missed something. So I use magic to run on the same design, the DRC or the manufacturing rules, the design rules manual, and the design rules themselves have to be the same. I can't just waive them. So we test against known outcome, which is we know a layout is violating, is by design is violating, violating specific rules. So if I see that when I run my check locally on the open source, I know that I got these. Now, that doesn't mean that you have every single error under sun is covered. And in this case, you start doing, adding more layouts that have different structures, different components, different types of transistors. And you just crunch them into the checker machine and then see what's clean is clean and what's not clean is not clean. And this process has a lot of good automation so that you can do it without actually having a manual intervention.

Chris Gammell: So it sounds like it would be really tough to do this if you were working with, I guess you have known designs. I mean, are these designs that have been checked against processes for years and years and years? Is that kind of the idea or is it, would it be really tough to do this as a leading edge kind of thing? Cause you wouldn't have as many like validated designs that work.

Mohamed Kassem: So actually for advanced, no, for, you know, let's say I'm back. I worked on, you know, 20 nanometers in 2010. That's a, back then it was the advanced quote unquote, a team that develops pass fail layouts and someone else is developing the checker. So they don't have the buffer, you know, they have an independent type of blindness. So you don't check for what you know. So the people do that until they figure it out. And sometimes they make mistakes and they end up with a bad silicon. Expensive, very expensive.

Chris Gammell: Yeah. Yeah. That'll, that'll, that'll ruin your Friday.

Mohamed Kassem: Yeah. Literally. I mean, if you find a bunch of metals that are supposed to be parallel, they're all one slab. Yeah. Yeah. Yeah. So, yeah. Yeah. So, uh, that's one side. Now, uh, the other side, which is going to be really cool is when it, because of the open source aspect of the Skywater PDK, you will find the number of layouts that are going to be generated. This is a public. So if I, people were actually building a place where you can store some of these in, in a, well, store a manifest of these in a way that a collection of layouts that are available. With known results are going to be available to people. So not just me can verify because I have access to the sort of manual through the boundary. I can have, I, you know, anyone else can do that and, um, have tons of designs that they can use for the first time, by the way.

Chris Gammell: Yeah. And that, and that will also allow people to make maybe some, you know, calculated risks for, you know, a, well, this hasn't been done before, but we ran all these other designs against, against stretching this one design rule. And now we can, we feel confident, you know, we feel 85% confident that it's going to work based on all of these designs and simulations that we ran against it. And we're willing to try it on this process or something like that.

Mohamed Kassem: Perfect. And, and there is a little, you know, when you're trying new, there is a little bit of judgment. And I, I personally, when in doubt, I go check with the process guys. Why did you have this rule? Why? Is it because something optical, like if you don't have this rule done right, it's going to have a short circuit? Or do you want it? I'll tell you, there's an amazing rule that it was about the width of the poly. So, uh, and the aspect ratio of the given poly on a transistor and 40 nanometer, 45 nanometer and on the transistor. So I was curious why. And he told me that when we do laser annealing, like annealing with a laser, so you heat it up. Like a fusing kind of thing after, after the end of the line or, or when? Well, it's on the poly. There is a heating process that makes the chemicals, you know, and the mechanical stresses just go settle in a certain way. So it turned out that when you do it with laser, it gets in and the poly acts like a wave guide. And then it hits. If you have the, the length of the poly close to the, the wavelength, it will actually become like a focused. And it burns. It burns. Wow.

Chris Gammell: That's crazy. That, that's, that's a, that's a rule that's been, that came from experience. That's a hard one rule. It sounds like. Yeah.

Mohamed Kassem: So I'm not going to discuss, I'm not going to negotiate that. Yeah. Right. Yeah. Yeah. So, so when in doubt, as I said, I go, I go, I go, go to the process, ask why the rule is, is designed this way. And then, cause it's always, sometimes they do it in a conservative way. Like they tell you, oh, well I can.

Chris Gammell: It's 40% greater than, than the last time they tried it or something like that. Or.

Mohamed Kassem: Yeah. Yeah. But keep in mind that there is one thing that if the process is stable, like something that's already been used for 10 years, you're very unlikely to have that happen. Because there's a lot of customers and a lot of users that are already used to certain rules. Right. If you change the rule on them, then you immediately invalidate their, whatever, depending on the rule, what the rule was. Right. So in this case, you justify the effort. You see, okay, the rule, if we change this rule, we're going to make the process more efficient. The tape out more efficient, for example.

Chris Gammell: Yeah. This, so this is interesting to me because I used to always wonder, like, you know, I talked to like TI, like factory application engineers who came through and they'd be talking about, oh, well, or, you know, we're moving to 180 nanometer. And this was like, you know, I guess probably 10 plus years ago when I was talking to them. And like, they were just, they were just getting to there as well. And I was like, well, you know, I'm reading news stories about 45 nanometer or whatever is the newest thing. Why are you guys on 180? And they're like, well, it's qualified for analog now. And I never really got it. I always just kind of figured it was just like a trailing edge thing. But it really sounds like now it's like the fact that it's validating over and over again, especially when you want to tweak things a little bit at a time, right? If you're doing an analog design and you want to change the gate width by a small amount or you want to do some kind of weird process step that is non-standard, you would want all these designs like you're talking about to validate it. And so it sounds like now having the open source stuff, it's just going to fast forward a lot of things because it's not just, you can't, I'm guessing here, but I'm guessing in the past you couldn't have gotten access to a lot of designs outside your company to go and validate a design kit. You'd have to sign a bunch of waivers or something.

Mohamed Kassem: And if someone didn't experience that collaboration intensity and the quality of outcome that the open source community delivers, you won't understand how this happens. It's crazy. You would think it's unreliable. You'd think it's not going to work. No, these guys are no politics, blunt. They figure out a problem. They tell you what it is. And sometimes they come with a solution as well. So having that is a great thing. Now you mentioned 180 is a failing thing. Any technology has a lifetime. You develop it and then you use it for many products and then you sunset it. So the 180 nanometer and 130, these things, they didn't get sunsets. This has continued forever because first of all, one of the things, if I try to iterate on some new architecture over and over with the $10,000 versus $100,000, iteration. So it's cheaper to actually iterate. And that's a part of the Google picture here is that continuous integration like this concept in the software. You put a chip, test it and learn and get another one in a lower cost.

Chris Gammell: Also, also the Google idea of throwing money at things. Yeah, that actually helps. Yeah. Hey, you know, we'll take it. Yeah. It's moving the industry forward. I'm good with it too. Yeah. Yeah.

Mohamed Kassem: And when you start thinking about it, the last problem I have right now, like today, I want more designs. My problem is not finding money for the fabs. Actually, the problem is in the design. The bottleneck is in the design. Right. So we need more designs. And sometimes because of this, the fabs would voluntarily say, oh, I like it. I'm going to just give you a shuttle. Sometimes. Okay.

Chris Gammell: Interesting. And why is that? Because they want to also test their process for something? Or what does that mean?

Mohamed Kassem: Once you have a design implemented on this process, it sticks. So it becomes a part of their portfolio offering. Ah, okay. Okay. So, so they say, I have this and it's still owned by you.

Chris Gammell: Right. But they keep a couple of wafers. They get all the failure analysis data and then they can. Yeah.

Mohamed Kassem: And they say, they say when a customer comes, if they come, they say, oh, we have this data converter developed by this company or this person. And that's an enablement. So more IP and more functions means potentially more business.

Chris Gammell: Interesting. So, so what I'm hearing here is that all of the people that are listening at foundries right now should think about this because they get free IP that can be easily validated. Not that we have like a lot of listeners that are in the foundry business, but that does, that is interesting as like a using, using the skywater thing as a, as a beachhead to try and get other foundries in the future and to be like, Hey, look, you guys are missing out on all this IP over here. Why don't you go and offer this now in the future and also maybe open up your process and make it more of an open ecosystem.

Mohamed Kassem: And I think that's started to happen. Uh, I mean, if you notice that the main announcement that happened was Tim Ansell's presentation and then the second presentation, uh, yesterday or two days ago, uh, now, and some Slack channel, but there is, there is an announcement coming from official, like they are from the foundry say, okay, we're doing this. And then when you have this, that's going to be, uh, even more assertive to the market saying a foundry has already come out and done this. Now the benefits, what's the rationale for the founder to do this? Not just the validation of IP or availability of IP. Here's the thing. Here's the thing. Statistically speaking, if you increase the number of designs, the chances of one of them to hit the jackpot in terms of market increases. If you only did two, you're really, really have to be careful to, you know, to try to make one of them or two of them really hit the market and make revenue. But when you have, you know, a hundred or a thousand, so the philosophy is very simple. I'm going to enable as many people as possible with open source tools and process. And then that increases the statistics of someone coming to me with the foundry. This is the foundry hat coming to me with a design that needs to be a product. Yeah.

Chris Gammell: Yeah. Cause at the end of the day, they, they, they, they just want to make as many chips as possible. And so, yeah, it is, it's basically just another, I mean, it's a marketing tool at a certain point. Right. But it's also, you know, there's benefits on the other side. So, uh, yeah. I call it a currency.

Mohamed Kassem: So it's almost like in order to build an IP or a block in a given technology, you're basically converting your money into that currency.

Chris Gammell: Like a itchy, itchy and scratchy bucks in, uh, in the Simpsons. Yeah.

Mohamed Kassem: Yeah. That currency has to be, if it's not used, then you lost your money. You just bought a wrong currency. Okay. That's right. Yep. So the foundries need to typically be in the process. Although they don't say it, but they, they, they say it in a way that my currency is good and it's fluid and people can use it. Right. And my chips would. Right. Now, when you see that people will come in and say, oh, I'm going to convert my money there or my time rather.

Chris Gammell: Yep. Right. And so you're saying in the past it was fiat currency and now it's a, it's a, it's a Bitcoin style. Yes. Yeah.

Mohamed Kassem: In a, in a sense. More, more fluid. The one last thing that happens here is that the foundry. Uh, so it's interesting that, uh, uh, there's a process that people always, when the foundry comes up with a new thing, the customers, the first customers become the guinea pigs. Okay. Yeah.

Mohamed Kassem: They, they, they, they, they, they try and the, and then over time, the foundry, when they sell to more customers, the quality of the output of the PDK, for example, increases because of the several, you know, loops of, of trials and errors. Yeah. Yeah.

Mohamed Kassem: Yeah. Yeah. That's a serial process to a certain extent. If it's a new thing, when you do it with the open source community, I mean, I saw an example for the Skywater, even though it's an already established process. We have in a couple of days, 200 people joined the Slack channel, at least 10 of them saying, I am here. I want to help test it.

Chris Gammell: Was this about two weeks ago? Because I mean, I'm going to take credit then.

Mohamed Kassem: Well, if it was two weeks ago, then it was the amp hour.

Chris Gammell: If it was before that, well, you know, okay. Then that was, I'm sure it was, yeah, it was the announcement. That's really great. Yeah. So you're saying that that basically enables just more parallel processing, basically.

Mohamed Kassem: And that is, that is amazing because statistics, again, statistics, and when you have parallel tracks, three paths to do the same problem, if each path has an 80% of hitting the problem and you do three, it's 99 per something percent. So, and then people, they, sometimes if they make them say, oh no, I found it, I made it work in this way. And they say their solution. And then these solutions become a documentation. It's a, it's a fascinating world. And I do think that, however, in order to post Google world, meaning when I say post Google, meaning in order for the semiconductor industry to change, you can't have just Google paying money to make free shuttles. Right. Right. Right. So there needs to be a change in the, in the dynamics, which we believe we've, we've been focused on it. As crazy as people say it, we've been focused on it to reduce the upfront cost to get to prototype and prove the design or prove your, your, your idea.

Chris Gammell: Yeah.

Mohamed Kassem: From then you can choose to go back up or down. And in this case, you have much more leverage. You have much more value in the, in the idea. And versus I'm, I'm, I have this idea that I will do if you give me the tools and the process. And, and, yeah.

Chris Gammell: Yeah.

Mohamed Kassem: So one, one, one thing that I didn't mention here, we've been talking about people who make chips, okay. Design chips, who know how to design chips and play with chips. Well, one of the objectives that we have, I mean, it's not rocket science. Okay. So it's kind of a shame basically to have the process of creating a chip since its inception. It hasn't changed.

Chris Gammell: Well, it's got a little smaller, but yeah, it's still, it's still, you know, P area and area or equivalent and, you know, mass process. Yeah.

Mohamed Kassem: But the intent, well, it changed obviously in the steps and, you know, there are details. However, it's a, there is designs, there is a modeling, there is create, you know, create the front end, back end, the creating GDS, putting your camera. Well, the, the one thing is that when you listen to this process requires a lot of knowledge. Well, I think in what, that's what we're doing in how this is that we need to simplify to a level where I don't need to know how it's made. I just want to use it. And I'm, and I tell you what to do. So like, so if I were building a board and if you go to the, you have this website right now, just on the landing and it's going to sound like a marketing, but it is in a sense, it's a good thing. So I'm going to say it. If you, if you go to the main website, you're going to find a template in front of you. It's called Raptor.

Chris Gammell: Yep.

Mohamed Kassem: And when you click on Raptor, this is an arm based chip that already exists and I have a board for it. And if you click more, it asks you, tell me what peripherals you want in a point and click.

Chris Gammell: Yeah.

Mohamed Kassem: And you add. So this is an interface. Who's using that interface? Not the chip designer. Is a system designer. Who says, I just need that much RAM. It's like a selector guide. I need that much RAM. I need two ADCs. I need that, that, that, that. Right.

Chris Gammell: This is basically a, it's a, it's a part family. It's like looking at a part family that has 80, I don't know how many variations there are, but basically it's every variation instead of like, like, like having to look at the table from like a microcontroller family, like an ST30, STM32 and saying, well, there's, you know, all these different options, but you only get like eight total options or maybe 64 total options of all the combinations you have there. There's as many options as, as the combination can have with, with this.

Mohamed Kassem: Yeah. And the more people add. So for example, some luck comes up with an accelerator. You can add it there on their ownership. So now if the system developer uses that accelerator, that validates it. And it's, there is actually a business relationship that actually, even if it's open source, the fact that that accelerator is used as a, you know, and served by the, supported by the developer. So it's a value for the receiver, but the language here of the, this is abstraction, meaning that I don't need to know. I don't want to hear GDS. I don't want to, I just want to tell you what I want in the chip and you. Right. This is, this is, this is a menu ordering system. Basically this is. Exactly. Yeah. And the, you know, the numbers for it, you'll find on the main website. If you, if you stick to that template and it's a template, we call it a template and, and, and the list of options in the menu, you're guaranteed to get your product in chips in a 70k, 70k dollars. This is a person, this is a company that's building, not, not, not, this is someone who's receiving the chip, not developing the chip. Okay. Developing the chip is much less because they apply their own experience. Receiving the chip, this number is probably at least 5x. 5x less you mean? Or 5x less? Less.

Chris Gammell: Sorry. Okay.

Mohamed Kassem: Yeah. Less. Okay. So when you say it's $70,000 to get the first hundred parts of my product, that is the, this is a very low entry barrier. And, and that's for the recipient, as I said, for the recipient and the, on the other side. And, and typically when you, when you go, it's becomes started at 200, 200, $300,000 to, to. Now, why is that number low? It's not because the chip is smaller or, you know, maybe it is smaller, but because it's templates are pre-engineered. So some of the engineering time that you already spent, it's already in it. And the changes that you're making are restricted to controlled changes. Like you put them on the bus, like the old computers where you slide the card into the, the back of the computer. That's right.

Chris Gammell: Yeah.

Mohamed Kassem: ISA slots and stuff. Exactly that. Exactly that. So we have a bus and just hook up the things to it. And then you, the more you add, calculate the area, calculate power, and then, so if you're doing that, you're guaranteed that the motherboard, you can fix the motherboard. You don't have to verify the motherboard every time you have to verify software, but not the motherboard. So that's the equivalent.

Chris Gammell: Okay. So I hope you, I hope you don't think I was listening, but I was clicking around as I was doing that. And in the amount of time from when you started talking about that to now, I just designed a chip. And so like, if people want to know, I was just clicking, I was pretty much clicking, you know, just, just as I saw a click and I did a click, but like, that's about how long it takes. So that was about a four minute process to, to make a complete, not a completely custom, but like, I just configured a chip that is, that is really cool. I mean, like.

Mohamed Kassem: And you can save it. And if you notice, you're going to find the cost.

Chris Gammell: Yeah. So it's a dollar 62 part cost. Yeah. Die area is 5.35 millimeters. Total IO is 4.4, 44. And then licensing costs $0. Yeah.

Mohamed Kassem: Because, because the IPs are on that chip specifically. They're not, uh, they're presented as a part of the package. It's not presented. Uh-huh. Now someone can put, uh, you know, uh, you can put their IP and say, okay, it's going to be X cents per part, whatever. And that shows up in the calculation. And so that gives you information about decision. You can actually save multiple configurations. You can configure it A, B and C and D. And then now you'd have multiple options with different costs, different trade-offs. So it's a very powerful tool. And then now just to complete the, the, the dream. Once you do that, there are two ways of getting the chip finished. You either someone of, uh, experienced in making chips, they will take your RTL. Because by the way, whatever you did on the form, it goes to become an RTL. So it's not, it gets compiled into a code, very long code that is ready to go to sound, to backend, to, to, to, to be hardened to the layout. Now we're doing that with open lane.

Chris Gammell: Okay. So that it's now configurable on your computer instead of with. Yeah.

Mohamed Kassem: And by the end, by the end of the clicks you had, you can literally click and wait for the right number of minutes or hours, depending on what the size of the chip is. And like strive would take a couple hours.

Chris Gammell: Okay. Yep. And when you say a couple hours, is it like, so there's like a flow and like, so we will link to the, the actual open lane. So there's a GitHub page for it and, and all that, but it, is it basically just like a make command? I mean, what is, what is it to actually kick off the process?

Mohamed Kassem: So if you're using the flow on our platform for a system, there are people who just want to get the chip. You don't want to deal with the flow, right? In this case, the, the, the last step of your form where you said, click, and here's my configuration. It gets passed automatically to farm that takes it, pass it through. So it's a equivalent to the make file, a sequence of steps. Each one of them has an outcome and then it runs all of them. And it gives, and the end of the last one is the GDS, but that starts from synthesis. And then you store the synthesis output so that you can verify it. And then your placement routing, DFT, et cetera. And you can sort of now it gets stored. But if I, again, if I'm the person who doesn't care about, I just want the chip. I don't care about that. It's, it's hidden inside the cloud. Now, if I am doing it the open source way, when you clone the open lane, it's a make. So you, or basically it's a, you can, you can do multiple layers. I mean, we, we did it in a Docker container. So to make it easy. And then the, another way to do it completely native, but basically it's a sick, you put the chip, you give me the design, you give me control files and then run and leave it.

Chris Gammell: Huh.

Mohamed Kassem: And then, by the way, this is an important part of open lane. I'm sorry to, is that now it's, what if it didn't finish? You know? So I liked Tim Ansel's expression and I give him that coin, a coin deter. I don't, one, one design rule violation is like a thousand. I don't know what to do with it.

Chris Gammell: Yeah. Yeah.

Mohamed Kassem: So, so the goal with open lane is to get you completely clean, ready to go.

Chris Gammell: Yeah.

Mohamed Kassem: Well, do you have to, in order to do that, the first configuration you put in terms of backend, how their floor plan, how their hierarchy, the hierarchy is managed, et cetera. It may not get you that first round. Yeah.

Speaker ?: Yeah.

Mohamed Kassem: So you change the configuration. Well, that, you keep doing that. Well, you could do that manually, but the flow had a picture like this. It says, try configure, we have a set of configurations, A, B, C, D, and then try them one by one and see which one gets you to the zero tier ready to go. Yep. And, and, and then now just to, to calibrate people because it's, you know, the way it's done, I'm not going to say I'm benchmarking it against, you know, any other commercial tool. I say it's the best benchmark because this is the only open source way to do it. That's one. So there's no other.

Chris Gammell: Take what you get. Yep.

Mohamed Kassem: The second one is that you sacrifice area, you sacrifice other things, but you don't have to be knowledgeable of chips to be, to get it. Yeah. So that's, that's, that's the key part.

Chris Gammell: Right. And, and it's interesting too, because I mean, I think one thing I think about with this is the, the idea of like adding, adding exactly what you need, or maybe it's a, you know, you, you just couldn't get the exact part you needed with an off the shelf part. Now you actually have an option. Yeah.

Mohamed Kassem: I have a friend of mine that calls this the right size compute or right size thing for what I want to do. Right. Rather than getting stuck with the standard. And of course there are standard parts that are really fitting what you're doing. Yeah. But, you know, deviation is deviation. You know, especially when you come to form factor power, memory, things like that.

Chris Gammell: Yeah. I mean, like as a practical example, I mean, like it would be really hard. It would, it would be really hard to buy a off the shelf microcontroller that had 16 UARTs, but it does seem like it's possible here. Is that, is that correct?

Mohamed Kassem: Yes.

Chris Gammell: Okay. That's pretty crazy.

Mohamed Kassem: And, and, and we adapt, for example, in order to put peripherals on the bus, the bus needs to adapt in terms of loading and other things. This is all happening in the background. Yeah. Yeah. And so exactly that is because, and no one is going to tell you why. Right. If that's what you want. If that's what, that's what you want to do.

Chris Gammell: Yeah. Right. Yeah. Yeah. And it's, you're not like talking to every vendor and asking like, well, can you, do you have, do you have 16 UARTs? And I'm like, no, why would we have that? We need to sell to more than one person. We need to sell, you know, we have to try and like hit the, it's like the Venn diagram. You have to try and hit as many people as you can. And just that one dot that's outside the Venn diagram, you're not going to make a custom chip for that person. But sometimes, sometimes the weirdo needs 16 UARTs, you know?

Mohamed Kassem: Yeah. Yeah. And I ordered ADCs or amplifiers or something. Right. Yeah. And you could prototype it some other way. And then, and that's what happened with one of our customers, actually. They did, like, did a board to prototype the whole thing. And he said, well, half of that board needs to go on a chip.

Chris Gammell: So how, I mean, how can people get started? What does it take to actually get to that point? Is it mostly just click like I did and register and then.

Mohamed Kassem: So there are two places. If you need, if you're, if you're passionate about developing and understanding how it works, you go open lane.io.

Chris Gammell: Okay. Okay.

Mohamed Kassem: It'll take you to the right place in terms of GitHub. And it has instructions of basically like systematic instructions or executable make. Where you clone and run it on your platform. And you will, it comes up with, it comes with about 40 examples inside it. So you can run any one of them. The process has inside the open lane instructions, how to get access to the, how to bring in the Google PDK in as well, because you choose what's library. All of that is included there. And then you, you, if you have a, an RTL code, very long, you can certainly use it and get your GDS right there. That's the, that's it. If you are, if you want to get your hands in the, in the down there and basically dirty with the flow and getting, getting in the, in the plumbing. If you want a chip and you, I don't care to run no open lane. I don't want to hear about open lane. All I want is that this chip that has 16 new arts. Okay.

Chris Gammell: Someone listening right now to like, oh my God, I have been waiting. Yeah.

Mohamed Kassem: And, and of course there's, you know, that, that, like the company that I mentioned is half the chip. They actually calculated it and found that the ASIC will actually make them make more money in terms of margin for, for that product, for their product.

Chris Gammell: That's great. Yeah. There's no, there's, there, there's no better way to like prove to your boss that you need to be doing something like this than to be like, oh, hey, money. Look at the dollars, dollar, dollar sign, dollar, dollar bills, y'all.

Mohamed Kassem: You know, and, and I think that's the reality we live in. So as much as open sources is a great thing in the hardware world versus software, because of the, and the cost of materials and the actual failure, if you fail, if you don't, iteration, the cost of iterations, because failure is okay. You can, you can reiterate, right. But because of that cost, it creates, you know, where do I get this money? Because it's actually physical stuff. You know, real hardware, you know, real hardware, things in your hand. So it's important to have a business model to continue to draw and keep that alive. It doesn't mean you sell your open source product. It becomes maybe something like a maybe in some areas becomes like the closest example is I developed an, an app or a stack or so quote unquote design. And then a customer comes and says, I need that design. You customize it for me.

Chris Gammell: Yep. Yep.

Mohamed Kassem: And then now, because you, you as an engineer, the developer, the earning there becomes, I made my design work for your system. Yeah. But the design is open source and the customer doesn't know how to deal with it.

Mohamed Kassem: Yep.

Chris Gammell: Yeah. Yeah. Yeah. Yeah. They, they basically, the open source was the marketing that allowed them to see it and see that you have the, the capabilities to make this thing happen. Actually, we had Ryan Cousins who was on the show. He talked about that with the board snickerdoodle that they made. It's like this FPGA development board. He said, the main part of their business is not selling the board. It's selling the customization for that board because people find out about it. And then they say, well, I don't want to deal with it. You, you do it. And then it's basically the exact same thing.

Mohamed Kassem: Well said. This is, this is exactly what it is. And we had the first, we had a, a way from a being, you know, just as a reality. When we were published Raven on GitHub, we got contacted by NEC research. And he said, Raven is good. We can do certain things with it. Give us some boards. We give them some boards. However, very quickly came back and say, well, I need that much memory. I need NBRAM. I need this, this, this, this. And we created Ravena. So you'd find a chip called Ravena on the platform. That's a, that's a customized version of Raven based on a customer input.

Chris Gammell: That's great. That's great. And it's, and then that comes back and that's another design that can be used again. And yeah.

Mohamed Kassem: And we have it with the board and everything ready to go. So people can, we send it, send these boards to the customer so they can actually develop and write, you know, test it for their own prototype at the system level.

Chris Gammell: That's super cool. Well, Mohamed, this has been really great insight into this. Where can people find out more about, about you specifically, but then also the company?

Mohamed Kassem: So Twitter, MKCosm, LinkedIn, MKCosm, the company, I'm on the website of the company and I'll be, I'm for sure that I'm, I'm inside the Slack channel or the Slack space for the Open PDK. And we're, and I want to share to, with people is that we're going to create a signup form that is a signup, a self-invite. So I don't have, people don't, you just go there to that place. It's going to be called join.skywater.tools. Join.skywater.tools. It's not ready yet. It will be that. Okay. People will sign up to the Slack channel space and they'll, they just get into that.

Chris Gammell: Right. Instead of having to ask for an invite or use the invite link that, yeah. I'm not sure why Slack did that. It's like one other, that's a, yeah. Anyways, Slack is its own thing.

Mohamed Kassem: Well, you know, whether you hit or like it, it is right now, it is the actual, right now we're benefit. I mean, the group of people, we have 300 people right now in that space. They're collaborating in multiple rooms and digital, you know, this, this, this, this, and it's working. But GitHub is always in the backend because if something has a bug, somebody has a bug or an issue or improvement, they file it as a GitHub. And so GitHub is, you know, is also, but that's an, and about the company, you go to eFabless.com and register. It takes you 10 minutes. You don't need an NDA or to pay to access any of these resources to design your own.

Chris Gammell: Right. You don't even need skill. Like I just showed, I mean, I just clicked and I made a chip. So I'm going to go put chip designer on my resume now. So thank you for that.

Mohamed Kassem: Well, and that's actually how it should be, to be honest, because, you know, driving the car. You don't need to know how the car works to be able to drive it.

Chris Gammell: That's a good analogy. Yeah. Well, thank you for joining us today. I really appreciate it. And I'm looking forward to chatting again in the future. Well, thank you. Thank you for having me. Appreciate it. We'll see you next time.