All right. Hey, Vic. Today's conversation is super interesting. It's scale-up interconnects, but it's not copper. It's not optics. It's radio frequency over a plastic waveguide. This is fascinating. So, it's your conversation with Sean Park, CEO of Point2. How did you come across Sean and why did you want to have him on Semi Doped?
The discussion is always between copper and optics, but so much of the conversation around this has never centered around a third option. So I first came across Point2 Technologies in my Substack where I've written about them in quite some depth, and then I ran into Sean at OCP APAC and I stopped by their booth, and so we were like, okay, let's get into this and talk on the podcast.
It's so much more fun when we can delve into all the details. So, let's listen to the podcast and the interview with Sean, and then I want to jump on with you afterward and discuss maybe if this can actually ship and if there's any sort of practical realities to deploying this. So, without any further ado, let's get into the conversation.
Welcome to the Semi Doped podcast, where we have conversations with the best minds in the semiconductor industry. I'm Vik Sekar, and I worked in the semiconductor industry for about 20 years, mostly on the engineering side. I am the founder of Semi-Exponent, where I do independent semiconductor research focused mainly on AI and hardware infrastructure, and I also write Vik's newsletter on Substack, where I explain how technology works and why it matters for the industry. With me is Sean Park, who's the president and CEO of Point2 Technology. He's a seasoned executive with decades of experience in semiconductors, wireless, and networking. Throughout his career, he has held leadership positions in many companies you'll recognize immediately, including IDT, which later became Renesas, TeraSquare, Marvell Semiconductor. As well as a CEO and CTO and a founder at TeraSquare, Sean was responsible for leading the company's strategic direction and overseeing its day-to-day operations. He was also a director at Marvell, where he provided invaluable guidance and expertise to help the company achieve its goals. He holds a PhD in electrical engineering from the University of Washington and also attended Seoul National University. Welcome, Sean.
Thank you, Vik, for the introduction.
Ah, I'm glad you liked it. I'm glad you liked it. Yeah, we've spoken, you know, in the past at OCP APAC, where I stopped by your demo and we played around with all the tech you had there, and it's exciting to have you on the podcast so we can discuss all the cool stuff we talked about for the general audience, because not everybody can be at OCP APAC. So that's why we have you here.
Thank you. Thank you so much.
Yeah. So, Sean, so basically Point2 has this very unique idea, and I like it because my own background is RF actually, and so I kind of relate to it closely. Everybody's always trying to connect data centers together with like either copper or optics. It's always like team red versus team green, or there's always like two sides fighting a battle. But you have this third approach, and you were like, let's send RF over a plastic cable. Like, tell us, like, basically, how did you come across this idea and decide you want to go down this path instead of everything else?
Yeah. So I have been involved in this data center market for the last 20 years, and previous to that, 20 years ago, I also worked on wireless systems, wireless chips. And in fact, the wireless signals, we call radio frequency signal, and that is really good for wireless systems, like a signal over air, and then connect two different or many different systems all together. And what if we use that RF signal over some sort of wired line communication, and then we can, in fact, we can extend the reach of cables in terms of distance, and also we can make a very simple silicon solution, and distance will be longer, and power will be much more efficient than existing optical systems. So that's the kind of concept that we studied.
That's nice. So very nice points you hit on here. One of them is, I think, when you say radio frequency, everybody imagines it's some kind of a wireless link. So I think one thing we want to clarify up front is like this is not a wireless link between two connections in a data center or anything. So this is actually a wired connection, except that the wire itself is a plastic tube or a plastic cable, and we'll get into why that is important and what the benefits are, and we'll also get into the points that you mentioned about how it saves energy in the data center and why that's like really important. That's a very good problem that you have in a different approach. So it's really nice to see innovation in different aspects. But this isn't a problem you've been working on like last year or two years, right? You've worked on this problem for a long time. Point2 has been around for a while now. So tell us about that journey.
So we spent many years researching the waveguide concept, and like as mentioned, RF over plastic concept is something that we really created first time, I believe. And then it's not really easy because, I mean, while the RF signal travels along some sort of plastic waveguide, in fact, you will see some dispersion and also some distortion while it travels through the cable. So that's kind of the fundamental problem we need to solve first. So we spent many years on how we can make the cable waveguide itself a linear medium so that we can transmit data over plastic, but when we receive it, it still maintains the same shape, same quality of the signal at the receiver part.
Yeah. So it's not as easy as people imagine that you can just send radio frequency over a plastic waveguide. We call it like a cable, but it's actually what it is is a waveguide. And waveguides have been around since like World War II, and they have been typically the guiding systems of radio communications for a very long time.
Exactly. Like the radar system. They put the waveguide pipe on the dish, and then they transmit very high power through the dish and put it in the air so that they can detect some object in the air. So that is the first concept of a waveguide: it's a metal pipe delivering high-power wireless signal to the dish. And then we use the same concept, but it's a flexible cable, and it's not really a metal pipe. So a metal pipe is kind of a defined boundary condition material, and instead the cable itself is very dynamic. It moves around, it can be bent. So it's not easy to put some signal in the flexible, dynamic cable and then receive that on the other side and still have a very clean signal at the end.
Yeah. So one of the benefits of the typical waveguiding structures where RF flows through is that it's actually surrounded by metal on all sides. This cable, and you have a sample which we will see later, is definitely not having any of this like metal over all sides kind of thing. Anyway, it's a different structure, and engineering RF to flow over this was the big challenge to get it to work right. But right now you're still a private company, but recently I was looking at all your financing and all that. It seems like there's a lot of interest. Do you want to fill us in on who your investors are, like who's really excited about your technology at the moment?
So anyone who sees the problem in scale-up network interconnect, they like our solution. The reason is that our solution is silicon plus a simple connector and a waveguide cable. It's very simple, cost-effective, low power, and easy to build. Complexity is a lot less than optics. So they like our solution. So for example, GPU companies, NVIDIA, ARM, they invest in our company, and also a foundry company in Taiwan, UMC, they invest in our company, and cable companies like Molex, they also invest in our company. So that kind of proves that our solution is very attractive in this market and in our ecosystem partners.
That's exciting. So like scale-up, the moment you say scale-up, I'm sure like half the listeners are going to prick up their ears here because it's actually kind of a problem in our current data center world. And I can imagine that this problem has emerged into your universe in the last couple of years because our GPUs are exploding in the data center and we need to hook up all these GPUs in a rack. Now, clearly when you mention scale-up, I noted that you don't mention scale-out. So by saying scale-up, immediately you're like, okay, we have a limited reach in which we want this technology to work, right? So let's set that expectation clear up front, like what is the kind of reach you're going for with this technology, and then we'll go into really how this thing works because we always cover the physics behind technologies here. So what's the reach?
Right. So our first generation will target 5 to 10 meters. That is a good distance for scale-up network. So within the rack, connect many GPUs together, and rack-to-rack applications.
Okay. So that's perfect. Because even when you say rack-to-rack, it's not really scale-out still nowadays, because the scale-up itself has so many GPUs you can't fit them in a rack. You have to put them in multiple racks, and now you have to connect them up as if they were one rack, for which you need cables. And so 5 to 10 meters covers even a multiple-rack scenario, right?
That's right, yeah. 5 meters we can connect four, and 10 meters we can connect eight racks. So that is good distance. So copper cannot reach.
Yes, copper cannot reach this distance, and that's why we need something else.
This episode of Semi Doped is presented by Crusoe. They build the power the data centers in the cloud behind AI. Did you know Crusoe can help you tune an open-source model on your own data for just a few bucks? Frontier models are very good, but they don't know your company's terminology, your part numbers, or how you like your reports formatted. So you keep stuffing instructions and context into the prompt of an expensive frontier model to get the job done. But for simple repeatable work like reports, an open-source model trained on your data can actually match a much bigger model. Think of a new grad with three months of on-the-job training. You don't always need the PhD. And the open model is cheaper and faster to run. Fine-tuning used to mean renting GPUs, working on orchestration and Kubernetes, writing training code, and babysitting the entire setup process. But Crusoe's serverless fine-tuning removes all of that. Upload your datasets, pick an open model like GLM, Llama, or DeepSeek, and they handle the training. You pay per million tokens used, not per GPU hour. And on the smaller models like Qwen 3 8B, $2 can take you a long way with millions of training tokens available. Once the training is done, you can then deploy them on Crusoe Cloud in just one click, or download the weights and run them anywhere. Your data only fine-tunes your model. And it doesn't have to be work data either. You can try it on how you write emails. Pick a task, compare the tuned model to your current prompt, and see how it does. Crusoe is giving all new accounts $5 in credit to get started today. That's enough to start tinkering with fine-tuning right now. Links in the description.
All right. Awesome. Now, we have to go to really understanding like how this works. So guide us through the whole process of the technology and, you know, how it works. Don't worry about the detail. We'll get into as much of it as we can. It's very exciting under the hood. So we have basically a simple problem. We have an electrical signal at one GPU and another electrical signal that has to be received by the GPU eight racks away. Let's say 10 meters, eight racks away. Now what is the process of signal conversion between the electrical here to the electrical signal eight racks away? What happens here?
So we have four different technologies inside our E2 platform. So first we need an RF transmitter. For example, on one side of the rack, let's say we have an RF transmitter, and that converts the baseband electrical signal into an RF domain signal, for example like 100 GHz, 200 GHz range of RF signals. And then since we're using an RF signal, we need to launch it into the cable, and for that part we need an antenna and we need an RF connector to couple the antenna and the cable. And then once it transmits through the cable part, that is the waveguide, and the other side is exactly the same components: connector, antenna, but it is a receiver. So all those components, especially the silicon part, is a fully integrated silicon with antenna inside the package, and the connector is a simple waveguide RF connector, and the cable is similar to copper. It looks like copper, but inside we don't have a copper core. So that's the main difference, and that's why we can still manufacture that cable using existing copper equipment.
Awesome. Can you show us the cable now? If you're watching on YouTube, you're going to benefit because Sean's going to show us the cable.
So I'm going to show you first the plastic core, that is a pure plastic material, very flexible. This is polyethylene.
And then since we are going to bundle several cores of this into one cable bundle, so we need to shield it. So we cover this white dielectric waveguide core, and then we put metal outside. This is for shielding purposes. So this cable looks exactly the same as a copper cable, but since we're using a wireless signal, an RF signal, we don't have a differential pair. We only have a single core, and then we also put multi-channel inside this cable core, so we can reduce the number of cables compared to copper.
Okay. So when you say that you don't have a differential pair, do you mean you will have two of these cables, these blue cables you showed us, for every connection?
Oh no, what I said is that we don't need a differential pair. Copper, they needed a differential pair, but we don't. So effectively we can reduce the number of cables by 4x compared to copper because we put two channels into one cable core, and then we don't have a differential pair. So roughly it's 4x better in terms of cable density.
Okay, that's very interesting. So when you say two channels, what exactly is that? Are there different frequencies that go inside?
Exactly. I mean, if you think about like FM radio or AM radio, you can use many different channels to deliver some data over the air. So we use the same concept. We use two channels, different frequency bands, and then we can put two data rates into one cable core.
Okay. So does that double the data rate across a single cable because of that?
That's right. So for example, this is a 1.6 OAP cable. 1.6 OAP, the IT standard is 200 GB per second per channel. But in our cable we are putting 400 GB per second per cable core.
Okay. So that cable already... two times 200. Okay. So each frequency that's within the cable, let's say 100 GHz like you mentioned, is 100 GHz carrying 200 Gbit per second, the 200 GHz is carrying 200 Gbit per second. So totally you have 400 gig.
That's right. I mean, in terms of data, this data rate is 400 GB per second, but in terms of frequency band we are using 90 GHz to 225 GHz. So it's less than, of course, it's less than 4 GHz of bandwidth because we are using PAM4 and then we are using up-conversion in the RF domain. So in terms of bandwidth requirement, it's a lot less than the data rate that you talked about.
Okay. So this is very interesting. So now clearly there is some part of the circuit, some transmitter, receiver, transceiver circuit that is converting electrical signals into these 90 and 225 GHz signals. Right. Do you have a sample of that and can you walk us through what that looks like and what it does?
Yep. So I already showed the cable parts. So this is our 1.6 OAP module sample. So if I open up the shell... then I'm showing you two pieces. One is silicon, and the other one is a connector on top of that silicon. So I will open up this connector, and then you will see transmitter on one side, the other side you will see receiver. Middle is the antenna.
Middle is the antenna. Okay. I'm going to describe what I'm looking at. So if people are listening to this, they'll know what I am seeing. So basically there are two, you know, this is like an OSFP connector that is kind of a standard form factor that goes into data centers, and inside it is like a PCB, and there are two chips inside, and one is a transmitter chip and one is a receiver chip, and on top of one of them was like a connector that Sean just removed to show us, and on either side is actually an antenna that is actually integrated into the silicon so that it can launch RF signals from the silicon into the plastic cable. Is that right?
Okay. Amazing. That looks really nice. So that's a nice thing about this, and I'm looking at this OSFP connector, and I'm like, wait, there is no laser inside. There's nothing like that's fancy. Oh, that's the connector. Okay. So that's the connector. Aha. Okay.
It's a very simple connector. So the bottom side goes on top of the chip and antenna, and the other side we have four slots. And then we can slot in this plastic cable like this. So very simple coupling.
Okay. Okay, I see. So what happens is that the underside of that connector sits on top of the chip where there's an antenna. So it will just couple all the signals from the IC into the connector. But the connector has a cable coming in from the side. So eventually the RF signal will get coupled into the cable, and that's how you get the signals converted from the electrical domain to the RF domain and sent over 90 and 225 GHz on this cable.
Yeah. What is fancy about it, I was mentioning this just now, this is very simple inside. Like if you see the OSFP modules of an optical module, for example, you'll see so many things inside it. You will see a laser, you'll see a driver, you'll see wire bonds, you'll see this, you'll see that. This looks very simple in comparison.
Right, very simple and fully integrated. Like I said, it's a fully integrated silicon solution.
Okay, silicon. Silicon is good. I think the industry likes silicon for one reason: it's well proven, it's tested. It's probably the most tested and most used material on planet Earth, and it is not sensitive to a lot of temperature issues like laser materials are, and so that's a big concern in data centers, right?
Right, so no laser, more reliable. Silicon, like I said, is proven. And also another good thing about silicon is this is very scalable in terms of volume. So we are using 12-inch wafers. I mean, demand in this type of interconnect is huge. The market is really, really big, and then it'll go even bigger. And then scalability in terms of production and volume is very critical. Our supply chain is based on fully integrated silicon, 12-inch, plus the existing cable manufacturing equipment like copper equipment, and a very simple connector. So in terms of supply chain, it's a lot simpler than any existing interconnect solutions.
Also this is not like a 3-nanometer technology which is like everybody is competing for capacity, right? Like, what is the... can you say the technology node used here?
Okay, okay. So this is very different from all the high-competition, you know, CPU, GPU-based, you know, that you can't get allocation for. You can get allocation for 55 nanometer, no problem.
No problem. And the tape-out cost is, I don't know if I can say that, but it's 50 times less than any other GPU tape-out cost.
Yeah, the tape-out cost is so much lesser that you can easily reduce the cost of the connector itself. So you don't have to deal with leading-edge silicon capacity, getting all that packaging, and none of that complexity is here. And do you have any... you said you use PAM4 already, but is there any kind of DSP anywhere in this system, or is it a no-DSP system completely?
So this is a non-DSP system. So we only need the RF transceiver. And then the RF transceiver, we have also baseband circuitries that handle the interface between, for example, switch ASIC or GPU to our chipset and module. So we have sort of an equalizer inside.
Okay. So you do some analog equalization, and it's not like full digital retiming or anything.
Okay. Okay. So it's just some light amount of analog equalization to recover the signal, and I bet that helps increase the bit error rate as you go to these higher speeds.
That's nice. That's very, very nice actually. So the other thing I was thinking about is when I'm looking at that module, you said the supply chain is really not an issue on the silicon side. I completely agree. The rest of it in the OSFP, just like PCB-based connectors, all this is fine. What about the cable itself? Is that difficult to get, or is that widely available?
So the cable itself, I mean, if you think about existing copper cable, so if you peel off layer by layer, starting from the middle, it has a copper core surrounded by an insulator, and then the very outer layer is another copper or aluminum shield or ground layer. So that structure is very similar to our structure. The only difference is in the middle we don't have a copper core.
And the insulation layer that's been used for copper cable, in fact, that is dielectric.
So we are using exactly the same material that has been used for copper cable. So extrusion of those dielectric insulator and plastic, the process is very similar to existing copper cable manufacturing.
Okay. So in the scale-up domain, this is going to be a lot of cable requirement here. So clearly we need to have enough of these plastic cables to be able to make it work, right?
That's right. That's right.
And one other thing I noticed when I actually held the cable that you showed us on the podcast here, when I was in person, is like it's really light. Yeah, it's really light. And couple that with the information you gave us earlier that this cable does the job of actually four copper cables in a sense. What immediately came to my mind is how much space and weight this cable will save in a scale-up domain, right? Like, can you explain in a sense why the weight and the density benefits matter for a plastic cable as compared to copper?
So density is very critical in scale-up. For example, NVLink 72, within the rack they have 72 GPUs, and then in order to connect 72 GPUs all together, they need an interconnect 72 times 72, that is 5,184 interconnects. So that is the number of copper cables in an NVLink 72 system. So a lot of cables, and the weight really matters, and also density matters because too many cables, too heavy, very difficult to handle, very difficult to assemble, and then the airflow is important for cooling, but too many cables of course block the airflow. And in our case, our first-generation E2 solution, we can reduce the volume of the cable by 4x, and then weight will be six times less than copper interconnects. So that really matters.
Yeah. I was trying to also calculate what you were saying. 72 times 72 is 5,184 GPUs, but that's only for NVLink 72. Like if we go to NVLink 144, you doubled it. So which means the number of cable requirements is going to go four times.
Exactly. Four times. That is more than 20,000 cables. It's not really sustainable using existing copper interconnects.
Right. Right. So that's why I think people were like, okay, I can't put any more copper in a rack, so let's go to multiple racks. So what you're saying is like if we can put 144 GPUs in a single rack and we can connect them all with plastic E2 cable, it's a good solution because for one, you can actually connect them all up if you have the cable density to be much lighter and more dense than what a copper interconnect offers in the same space.
You don't have to go to multiple racks if you can avoid it, which is the best solution.
Yeah. If it works, then of course it's the best solution. We will scale like silicon. Yeah. Yeah. Exactly. So since it's completely shielded, I'm guessing that just like copper, there's no crosstalk issues when you put all these cables in a rack, right? Like they're all shielded anyway, right?
Right. I mean, in our case, we really care about the shielding because we need to minimize the crosstalk and interference from inside to outside. So we use the same concept as copper. The outer layer is a copper shield, and then that kind of prevents all the interference. And then we measure the performance is better than 50 dBc. It's a really, really good number.
Nice. Nice. Yeah, this is nice because not only do you have something that is lighter than copper, but you have something that is probably more flexible than an optical cable. Is that true? You think the bend radius is nice on this thing, plastic cables?
Yeah, let me show you. So we can, I mean, of course mechanically we can bend any shape, but in terms of electrical performance, our bending radius is 4.7 mm. So 4.7 mm is, I can still wrap it around my finger. That is roughly 5 mm. So this is really good bending radius.
So you just wrap it around your little finger, right? And you're like, okay, that's about 5 mm.
Yeah. My little finger is still bigger, so I think I need to use a smaller finger, but that's the smallest I have.
Yeah. Yeah. No, that's good. Like I think bend radius is important too, because you've got to be able to route these cables in a rack, and it's not like you can keep them all straight all the time. So that's also a consideration.
Yeah. So you know, like this cable industry is now splitting into like active electrical cable, then on the other side you have active optical cable. You know, like in the optical cable, nowadays everybody's talking about VCSELs or you're talking about microLED. And so what is your version of the cable called?
We call it active RF cable. In short, it is ARC or ARC.
