RF Over Plastic Is a Third Option Between Copper and Optics: Point2's Sean Park

0:00 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 Point Two. Uh, how did you come across Sean and why did you want to have him on semidote?

0:21 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 Point 2 technologies uh 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.

0:43 And so we're 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.

0:50 So let's listen to the podcast in 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 uh without any further ado, let's get into the conversation.

1:12 Welcome to the semi-dope podcast, where we have conversations with the best minds in the semiconductor industry. I'm Vic Shaker and I've 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.

1:28 And I also write Vic'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 Point 2 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 Renaissance, TerraSquare, Marvel Semiconductor as well.

1:58 As a CEO and CTO and the founder at TerraSquare, Sean was responsible for leading the company's strategic direction and overseeing its day-to-day operations. He was also a director at Marvel, 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.

2:21 Welcome, Sean, welcome to Semidoped.

2:23 Thank you, Vic for the introduction.

2:25 Ah, I'm glad you liked it. I'm glad you liked it. Yeah, we've spoken to, you know, in the past at OCP APAC, where like I stopped by your demo and we like played around with all the tech you had there. And it's exciting to have you on the podcast so we can like discuss all the cool stuff we talked about for the general audience because not everybody can be at OCP APAC.

2:50 So that's why we have you here.

2:53 Thank you. Thank you so much.

2:55 Yeah, so uh Sean, um so basically point two has this very unique idea, and I like it because my own background is RF actually, and so I can kind of like relate to it closely. Everybody is 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.

3:24 Like, tell us like basically how did you come across this idea and decided you want to go down this path of everything else?

3:31 Yeah, uh so I have been involved in this data center uh market for the last 20 years. And previous that 20 years ago, I also worked on the wireless wireless systems, wireless chips. And in fact, the wireless uh signals is recalled 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.

4:01 And what if we use that RF signal over some sort of wireline communication? And then we can if in fact that we can extend the reach of uh cables in terms of distance, and also we can make a very simple, uh simple silicon solution, and distance will be longer and power will be much more efficient than existing optical system.

4:28 So that's the kind of concept that we study.

4:31 That's nice. So very nice points you hit on here. Uh, 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.

4:55 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. Um that's a very good problem that you have in a different approach. So it's really nice to see innovation uh in in different aspects.

5:15 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. Point two has been around for a while now. So tell us about that journey.

5:28 So we spent many years on and researching on the uh waveguide concept. And like I mentioned, RF over plastic concept is something that we really created uh first time, I believe. And then it's not really easy because the I mean, while the signal, RF signal is is travel along the uh some sort of plastic waveguide.

5:55 I mean, in fact, uh it will you will see uh some dispersion and also some distortion while it travels through the cable. So that's kind of fundamental problem we need to solve first. So we spend many years how we can make the cable waveguide itself is linear, medium, so that we can transmit data over plastic, but when we received it, it still maintained the same shape, same quality of the signal uh at the receiver part.

6:28 Yeah, so it's not as easy as people imagine that just you can just like 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.

6:48 Exactly, like uh like uh the radar system. RADA system, they uh they put the waveguide pipe on the dish, and then they transmit very high power through the dish and the uh put in the air so that they can detect some object in the air. So that is the first concept of waveguide, it's metal pipe and delivering high power wireless signal to the dish.

7:16 And then we use the same concept, but it's flexible cable, and it's not really metal pipe. So metal pipe is kind of defined boundary condition uh material. And instead, the cable itself is very dynamic. It's move around, it's it can be banded. So it's not easy to put some signal in the uh flexible, dynamic uh cable and then receive that on the other side.

7:43 We still have very uh clean signal at the end.

7:48 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 uh like metal over all sides kind of thing, anyway. It's it's a different structure.

8:06 And engineering RF to flow over this was the big uh challenge to get it to work, right? Um but I right now you're still a private company, but recently I I was looking at like all your financing and all that. It seems like there's like a lot of interest. Do you want to fill us in like on who your investors are?

8:24 Like who's who's really excited about your technology at the moment?

8:27 So anyone who see the problem in scale-up network interconnect, they like our solution. The reason is that our solution is silicon plus simple connector and couple like a waveguide cable. It's very simple, cost-effective, low power, and easy build. Complexity is a lot less than uh optics. So they like our solution.

8:53 So, for example, like a GPU company's NVIDIA, ARM, they invest in our company. And also a foundry company in Taiwan, UMC, they invest in our company. And cable companies like a 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.

9:16 That's exciting. So, like scale up, the moment you say scale up, like 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 really problem has uh emerged into like 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.

9:41 Now, uh clearly, when you mention scale up, uh, I I I noted that you don't mention scale up, scale across. So by saying scale up, immediately you're like, okay, we have a limited reach in which do we want this technology to work, right? Um so let's set that like expectation clear up front. Like, what is the kind of reach you're going for with this technology?

10:01 And then we'll go into really how this thing works. Because we always cover the physics behind technologies here. So, but what's the reach? Right.

10:09 So our first generation ETube will target five to ten meters. That is a good distance for scale-up network. So it within the rack, connect many GPUs together and rack-to-rack applications.

10:24 Okay, so that's that's perfect. So 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.

10:41 And so five to ten meters covers even multiple rack scenario, right?

10:45 That's right. Yeah, five meters we can connect four, and ten meters we can connect eight racks. So that is good distance. So copper can then reach.

10:55 Yes, copper cannot reach this distance, and that's why we need something else.

10:59 This episode of Semi-Doped is presented by Crusoe. They build the power, the data centers, and 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.

11:21 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.

11:37 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 Jeepseek, and they handle the training.

12:03 You pay per million token used, not per GPU hour. And on the smaller models like Quen38B, $2 can take you a long way with millions of training tokens available. Once the training is done, you can then deploy them on Cruso Cloud in just one click or download the weights and run them anywhere. Your data only fine-tunes your model.

12:23 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's 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.

12:41 All right, awesome. Now we have to go to really understanding how this works. So so guide us through the whole process of the technology and uh you know how it works. Don't worry about the detail. We'll we'll get into as much of it as we can. It's very exciting under the hood. So we have basically we have a simple problem.

13:00 We have 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?

13:22 So we have a four different technologies inside our EQ platform. So first we need RF transmitter. For example, uh on one side of the rack, we let's say we have an RF transmitter, and that converts the baseband electrical signal into RF domain signal, for example, like 100 gigahertz, 200 gigahertz of range of RF signals.

13:48 And then, since we are using RF signal, we need to launch into cable. And for that part, we need antenna and we need RF connector to couple antenna and cable. And then once it uh transmits through the cable part, that is waveguide, and the other side is exactly the same component. Connector, antenna, but it is receiver.

14:13 So uh all those components, especially for silicon part, is a fully integrated silicon with antenna inside a package. And connector is simple waveguide RF connector. And 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.

14:41 Awesome. Can you show us the cable now? If you're watching on YouTube, you're going to benefit because Sean's gonna show us the cable.

14:47 Um I'm going to show you first plastic core that is a pure plastic material, very flexible. This is polyethylene. And then, since we are going to bundle the several core 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.

15:16 This is for shielding purpose. So this uh cable looks exactly the same as copper cable. But since we are using wireless signal and RFS signal, we don't have a differential pair. We only have a single core. And then we also put uh multi-channel inside this uh cable core, so we can reduce the number of cables compared to copper.

15:45 Okay. So when you say that um you will have a differential pair, do you do you mean you will have two of these cables, these blue cables you showed us uh for every connection?

15:58 Oh no. What I said is that we don't need differential pair. Copper, they need a twin X 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 differential pair. So roughly is a 4x better in terms of a cable density.

16:24 Okay, that's very interesting. Uh so when you say two channels, what what exactly is that? Are they different frequencies that go inside?

16:31 Exactly. I mean, if you think about like a FM radio or AM radio, you can use many different channels to uh to deliver the some data over the over over the air. So we use the same concept. We use the two channels, different frequency band, and then we can put two data rates into one cable core.

16:56 Okay, so does that double the data rate across a single cable because of that?

17:01 That's right. So for example, this is 1.60 cable. Uh 1.60, the IEEE standard is 200 gigabit per second per channel. But in our cable, we are putting uh 400 gigabit per second per cable core.

17:18 Okay, so that cable already has.

17:20 Two times 200.

17:21 2 times 200. Okay, so each frequency that's within the cable, uh, let's say 100 gigahertz, like you mentioned, uh, is 100 gigahertz is carrying 200 gigabit per second. The 200 gigahertz is carrying 200 gigabit per second. So totally you have 400 gig.

17:38 That's right. I mean, in terms of data, this uh data rate is uh 400 gigabit per second, but in terms of frequency band, we are using 90 gigahertz to 225 gigahertz. So it's less than, of course, it's less than four gigahertz of bandwidth because we are using 104 and then we are using uh up conversion in RFO domain.

18:00 So in terms of bandwidth, requirement is a lot less than the data rate that you talk about.

18:07 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 gigahertz signals, right? Like you have a sample of that, and can you walk us through what that looks like and what it does?

18:28 Yeah, so I already showed the cable parse. So this is um our 1.6T OSAP module sample. So if I open up the shell.

18:39 Oh, look at that, yeah.

18:41 And I'm showing you two pieces. One is silicone, okay, 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 one side, the other side you will see receiver.

18:58 Okay, so middle is antenna. Middle is the antenna. Okay, I'm gonna describe what I'm looking at. So if people are listening to this, they'll they'll know what uh they're uh what I am seeing. So basically there are you know, this is like an OSFP connector that is kind of a standard form factor that goes into data centers, and there are inside it is like a PCB, uh, and there are two chips inside.

19:20 And one is a transmitter chip and one is a receiver chip. And uh on top of one of them uh 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 uh plastic cable.

19:42 Is that right?

19:42 Exactly. Exactly.

19:44 Okay, amazing. Uh that looks really nice. So that's that's the nice thing about this is that when I'm looking at this OSFP connector, I'm like, wait, there is no laser inside, there's nothing like that's fancy. Oh, that's the connector. Okay.

19:59 That's the connector.

20:00 Uh-huh. Okay.

20:01 It's a very simple connector. So bottom side goes to the goes to on top of the chip and antenna. And the other side we have uh four slots.

20:14unknown Yeah.

20:14 And then we can slot in this plastic cable like this. Okay. So very simple coupling.

20:24 Okay, okay. I see. So what happens is that the 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 you know the signals converted from the electrical domain to the RF domain and sent over 90 and 225 gigahertz on this cable.

20:56 That's right. Exactly.

20:58unknown Yeah.

20:59 This is what is fancy about I 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 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.

21:16 This looks very simple in comparison. Right.

21:18 Very simple. And fully intrated, like I said, is a fully integrated silicon solution.

21:24 Okay. Silicon. Silicon is good. I think the industry likes silicon for one reason. Uh, it's well proven, it's tested. It's probably the most tested and most used material on on planet Earth. Um, you know, um 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?

21:48 Right. So no laser more reliable is silicon, like I said, is proven. And also another good thing about silicon is that this is very scalable in terms of volume. So we are using 12-inch wafer. I mean demand in this type of interconnect is huge. Market is really, really big, and then it'll it'll go even bigger.

22:12 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 equipments like a copper equipment and very simple connector. So in terms of supply chain, it's a lot simpler than any existing interconnect solutions.

22:35 Also, this is not uh like a three-nanometer technology, which is like everybody is competing for capacity, right? Like this, what is the can you say the technology node used here?

22:46 So this is 55 nanometer.

22:48 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.

23:00 No problem. And the tape out costs is, I don't know, I can say that, but it's 50 times less than any other GPU tap cost.

23:09 Yeah, it's it's uh the the tape out costs are like so much uh lesser that you can easily you know uh reduce the cost of the connector itself. So you don't have to uh deal with like leading a silicon capacity, you know, getting all that packaging and none of that complexity is here, right? Right. And do you have any you you said you use PAM for already, but um is there any kind of um you know like DSP anywhere in this system, or is it like a non no DSP system completely?

23:42 Oh, this is a non-DSP system. So we only need a RF transceiver, and then RF transceiver, we have also baseband circuitries that handles the interface between, for example, switch AZ or GPU to our our chipset and module. So we have sort of an equalizer inside.

24:03 Okay, so you do some analog equalization and uh it's not like full digital retiming or anything.

24:10 No, it's not.

24:11 Okay, so it's just some light amount of uh analog equalization to recover the signal, and I bet that helps increase the bit at a rate um as you you know go to these higher speeds.

24:24 That's right.

24:25 That's right. That's that's very very nice, actually. So the other thing I was thinking about is when I'm looking at that module, is um you said like the supply chain is really not an issue uh on the silicon side. I completely agree. The rest of it in the OSFP is just like PCB-based connectors, is it's like all this is fine.

24:44 Uh, what about the cable itself? Like, is that like the difficult to get or is that widely available?

24:50 So cable itself, I mean, if you look, if you think about existing copper cable, so if you peel off layer by layer, so starting from the middle, it has a copper core surrounded by insulator. And then out our very outer layer is the another copper or aluminum shield or ground layer. So that structure is very similar to our structure.

25:19 The only difference is in the middle we don't have copper core. And the insulator that's been used for copper cable, in fact, that is dielectric. So we are using exactly the same material has been used for copper cable. So extrusion of those uh dielectric insulator and plastic, the process is very similar to the existing copper cable manufacturing.

25:47 Okay, so in the scale-up domain, uh this is going to be a lot of cable requirement here. So, like clearly, we need to have enough of these plastic cables to be able to make it to 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.

26:09 Yeah, that it's really light. And you couple that with the information you gave us earlier that you know this actually uh is this cable does the job of actually four copper cables in a sense. Um what immediately came to my mind is uh how much space and weight this cable will save in a scale-up domain, right?

26:33 Like, can you explain, like in a sense, why the weight and the density benefits matter for a plastic cable as compared to copper?

26:43 So density is is very critical in scale-up. For example, MVLink 72. MVLink 72 it within the RAC, they have 72 GPUs. And then in order to connect 72 GPUs all together, they need uh interconnect, 72 times 72 interconnects. That is 5184 interconnects. So that is the number of couple cable in MV Link 72 system.

27:13 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 E-Tube solution, we can reduce the volume of the cable by 4x, and then weight will be six times less than the copper interconnects.

27:44 So that really uh matters.

27:47 Yeah, uh, I was trying to also calculate what you were saying. 72 times 72 is uh 5184 GPUs, but that's only for NV Link 72. Like if we go to NV Link 144, you doubled it. So which means the number of cable requirements is going to go four times.

28:06 Exactly, four times. That is more than 20,000 cables. It's not really sustainable using existing copper interconnects.

28:15 Right, right. Um, so that's why I think people were like, okay, like 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 cables in a single rack and we can connect them all with plastic uh e-tube cable, um it's 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 can interconnect offers in the same space.

28:44 So you don't have to go to multiple racks if you can avoid it, which is the best solution.

28:48 Yeah, if it works, then of course it's the best solution. It will scale like a silicon.

28:54 Yeah, yeah, exactly. How does this uh what 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 in a rack, right? Like they're all shielded anyway.

29:08 Right. I mean, in our case, we really care about the shielding because uh we need to minimize the crosstalk and interference outside from inside, inside to outside. Yeah, so we we use the same concept as copper. So outer layer is a copper shield, and then that kind of prevents the old interference, and then we measured performance is better than 50 dBc.

29:34 It's a really, really good number.

29:36 Nice, nice. Yeah, this is this is nice because um 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 cable?

29:51 Yeah, let me let me let me show you. So we can, I mean, of course, uh mechanically we can bend any any shape, but in terms of electrical performance, our bending radius is 4.7 millimeters. So 4.7 millimeters is I can still uh like a round it across my finger. That is roughly five millimeters. So this is really good bending radius.

30:20 So you you just wrap it around your little finger, right? And you're like, okay, that's about five millimeters.

30:25 My little finger is still bigger, so I think uh I need to use a smaller finger, but that's the smallest I have.

30:32 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 you know, it's not uh like you can keep them all straight all the time. So that's also a consideration.

30:44 Exactly.

30:46 Yeah. So you know, like this cable industry is not splitting into like active electrical cable. Then on the other side, you have active optical cable, you know, like in the optical cable, can nowadays everybody's talking about vixels or you're talking about micro LED. And so what is your version of the cable called?

31:05 Uh we call it active RF cable. In short, it is ARC or ARC.

31:12 ARC, ARC, okay. The ARC cables then will go um in the scale up domain, which is very important because I don't think anybody should compare this technology to scale out. Like this is not a replacement for scale out optics. You know, 50-meter reach, 100-meter reach, you know, that's something you cannot compete with.

31:30 I mean, optics does a great job in that.

31:33 Right. I mean, optics for scale out, it was always optics. I mean, in the in the past it was optics, and in the future it will be optics, that's for sure. So Denmark is growing. That is a good market, but that's optical solution market, not for us. I mean, we specifically intentionally target scale up.

31:55 It's very high volume, very important market.

31:59 Yeah, that's good. So that's very clear, like you know, to distinguish because many people I feel like they confuse technologies, and uh there is always a right technology for the application. You you have to find, you have to match it to the right technology because it's not always about reach, it's about cost, it is about uh uh scalability, you know, it's about how many errors you can correct for or use, it's about temperature stability, reliability, you know, there are so many variables that matter.

32:32 That's right.

32:32 That's right. So real reliability is really important for scale up and especially for UL link and VLink type of interconnect. So they try to put the interconnect inside a box, inside the system box, and then they want to put the interconnect right next to their XPUs, GPUs, and MPUs. And if they put interconnect solution, active interconnect solution inside a box, reliability really, really matters.

33:03 If something happened, you need to replace it. But optics, I mean, sorry to compare always with optics, but optics is relatively non-reliable solution compared to silicon. So it's kind of dangerous for them to use optical solutions inside their system.

33:23 Yeah, that's that's no, no, the in the optics domain as well, today you see a lot of talk about putting optics closer to the GPU and XPU, right? And you have solutions like uh CPO that comes up, and it's always a question of reliability there because, like you say, when you have uh interconnects that go right up to the GPU inside the rack case, it's not like on the external to the rack, it's inside.

33:50 It's very difficult to change anything if something goes wrong. Think about it like each rack has 5,000 cables. Like, how do you know what went wrong? Like, how do you know what to change? Now multiply that right by you know all the racks in a data center. It's a massive problem, right? But I suppose since you only have a silicon chip there as a transceiver that converts electrical to um RF signals, you could very well use the co-packaged approach with RF as well, right?

34:20 Exactly. So that that was what I was trying to say. I mean, so our first application will be a pluggable uh use case like an ARC, ATF RF cable. And the second use case is inside the system, inside the box. So that is uh like a backplane solution to connect the many GPS all together, scale up. And those cases uh uh basically we only need silicon, a lot more relevant than any other optical solutions.

34:53 So we plan to do that. We call it near packaged ETU or co-packaged ETU.

35:00 So that's NPE or CPE, if you want to.

35:03 That's right, NPE and CPE. Exactly.

35:06 Right. And one other nice thing maybe I should um ask you to like explain about at least what I think is nice, is the fact that do you need any kind of gearboxing chip? And for those listening, if you don't know what a gearboxing chip is, you know, a lot of the interconnect solutions today have uh a wide but slow approach, which means that you have to convert a 200 gig uh lane into you know four 50 gig lanes or you know um 45 gig lanes, right?

35:41 So you have to make that conversion because you can't launch into an interconnect that only runs either at 50 gig per second when your incoming signal is 200 gig per second. So there's a chip that sits in the middle called a gearboxing chip that converts 200 gigs to maybe 50 gigs. Like, but for your case, the nice thing that you have in the scale-up domain that things like microvixels don't have and micro LEDs don't have is that you don't need a gearboxing chip.

36:09 Is that right?

36:10 We don't we don't need gearbox chip at all. So our electrical in interface 100% comparable to existing or future I2P data rate. So we take a 200G data rate as it is, and we just convert that to RF domain and we deliver that signal over plastic cable.

36:30 Okay, that's very helpful. So because these gearbox chips are actually usually leading edge silicon chips, they're they're not 55 nanometer chips, right?

36:39 It's very expensive and very uh power hungry.

36:44 Okay, perfect. Power. Let's talk about power. It's a very important question, especially in the world of um you know uh data centers and scale up, because the one thing that has stopped optics uh into become prominent in the scale up domain, and Nvidia has said this at many events that even I have been to, they're like, no, it's too expensive, energy expensive.

37:08 It's not about cost. They're like, it sucks too much power. I cannot spend per rack anywhere between 20 and 40 kilowatts, okay, just for networking. To put this in context, in the cloud era of data centers, the entire rack with all the compute and everything in it used to only be 15 or 20 kilowatts a rack.

37:28 Now our data center racks have become like 200 kilowatts. And now we are talking about spending what used to be a whole rack worth of compute only for networking. And Nvidia has always said, like, no, we will stay with copper as long as we can because we can't spend this much energy on just in networking.

37:46unknown Yeah.

37:47 So what is the point?

37:50 Like, what does the eTube offer in that in that um aspect?

37:54 So our power efficiency is a lot better than optives. I mean, if you think about why optics consume a lot of power, because they need electrical optical conversion. And for that purpose, they need uh laser, photodetector, they need some interface silicon like a TIA and driver, and sometimes they need DSP.

38:16 In our case, we only need fully integrated silicon chips. So power efficiency is a lot better than optics. So we we always claim that it's 3x better. And for example, this 1.6T first generation, we only consume uh 10 watts, less than 10 watts per module. So that is in terms of efficiency, it's like a five pico joule per bit.

38:45 And for NPE and CPE use case, our power concern will be less than two picojoules per bit.

38:54 That's comparable to the CPO you know, co-packaged optics uh level of promised uh energy efficiency per bit, right? That's the same range.

39:05 Same range, but I think uh our solution in terms of power concern, it will scale as the silicon scales.

39:14 Which means you can go to a smaller node and you can save the transceiver power.

39:19 Exactly. I mean that's kind of fundamental benefit of all the silicon technology we we had, right? So in the past our for example, I have been working on the Wi-Fi, Bluetooth, those kind of wireless chip designs. And every year, and and Marvel, every year we reduce the power custom by 30-40 percent by migrating from one process node to the other process node.

39:49 I mean, it will definitely happen to us as well. So tomorrow we'll hit uh two picojou per bit, and the day after tomorrow we'll we'll hit one picoj per bit.

40:00 Yeah, so you'll go to maybe a 22 nanometers or something, and then yeah, that's right. Because it's still not a leading edge technology. At 22 nanometers, you're not even in the FinFET era of technology, you're still in planar MOSFETs.

40:14 Yeah, not at all. I mean, we are making use of the uh the wireless domain uh silicon technology. We don't really need fast digital circuitries at all. So that's kind of another benefit of our process.

40:29 The fact that you don't need digital circuitry or DSP or uh forward error correction, things like that to happen at the module level, it might still happen on the host level, uh, but it won't happen on the module because you don't have a DSP there. Straightaway, one advantage that comes by just pure analog equalization is that you have better latency.

40:52 That's right. In terms of latency, DSP versus our solution, the difference is 1000 times. So our latency is like a picosecond, nanosecond, I mean below nanosecond range.

41:08 Okay. Yeah, that's very important because scale up, uh in the scale up domain, especially during training runs when you have these all reduced operations, the training run becomes as slow as the slowest link in the network. So you need all of them to have the lowest latency possible. Um, and DSP usually doesn't allow this because to do the DSP, you need all the bits to arrive.

41:32 And once the bits arrive, you do some error calculation and recovery and all of that stuff. All of that takes too long and too much power. DSP suck power too.

41:41 Exactly.

41:42 You avoid all of that.

41:44 We don't we don't have that. We don't have that, and then license is a lot better. And then yeah.

41:51 So within the pluggable module itself that you showed us today, right, what is your biggest challenge within that module? Like clearly, some part of it has to be the hardest for you. What do you think is like your biggest challenge within it?

42:04 So, in fact, we we have spent many years on eTube technology. I mean, nothing was really easy. First, think about the frequency range that we are using. We are using like a hundred gigahertz, 200 gigahertz of range of signals, right? So nobody has done 200g circuit design before. So that was the first challenge that we faced before, but not anymore because we already have silicon.

42:33 And on top of that, we still need to build kind of our own ecosystem. I mean, relatively easier because we still we can still make use of existing copper manufacturing equipment and then ecosystem. But still, we need to convince cable manufacturers and we we need to convince uh other ecosystem partners so that they can move on.

42:58 So that was kind of a difficult part in the past.

43:02 That's you also because you have this uh near packaged and co-packaged e tube so uh ideas. Would you consider also being part of like the open CPX MSA? So this is like a consortium. This is a consortium of companies who are coming up with a solution to uh put uh it's an open ecosystem of providers for those listening, uh, so that multiple people can supply the equivalent of co-packaged whatever, co-packaged copper, co-packaged eTube, co-packaged optics, like you know, so you fit into that ecosystem, right?

43:37 In fact, we are part of the uh consolution.

43:40 Okay, that's great. Yeah, that's that's good because now we have so many different uh competing solutions in the co packaged space that I think yours would be another one of them that is uh interesting and uh you know provides more options to people looking to do scale up.

43:55 Exactly. We'll we will give freedom for system engineers to. Decide.

44:01 Yeah, yeah, choose the best. So that's I want to also cover in general. I think we've covered like seriously, like I think I like this technology in the sense that it's different. And I always like the third option that we like, okay, like why do we always have to stick to two technology? Somebody's doing something totally different.

44:19 I totally, totally love that. So, in terms of actually getting into the getting this stuff into the hands of a customer, like say a hyperscale or somebody building, like what would they need to happen to this technology so that it is currently now going to be used in data centers? Like, what do they want from you?

44:38 So, I mean, we have spent many years to convince uh our partners and also our customers, including hyperscalers, GPU companies, including Nvidia. And right now we are at the point to deliver our product in their hands so that they can test in their labs, their environment, and they they really need to use in their system.

45:04 So that's the next step. After that, we of course we need to go through the qualification process. That will take many months.

45:13 And then the qualification process itself, uh, do they have to do reliability testing? And what what is the steps in the qualification that usually are required for interconnects and scale up like this?

45:24 Right. I mean, all those uh qualification processes is very similar for uh eTube and also similar to uh copper and also the optics. But usually they they call it gorilla test. So they will kind of plug in and plug out and twist, and then they they will run a script to all different use cases for each equipment.

45:48 So I mean it's it's not easy process, but no matter what, we need to go through it.

45:54 Yeah, they will put your uh ETube cable to the test, right? Like they'll put it in temperature, they will uh give it mechanical abuse.

46:02 Exactly, mechanical abuse, and then some script to kind of try to find out the uh like breakpoints. So yeah.

46:11 Yeah, that's that's interesting. So this is what it takes to get uh you know, this is useful because I think many other startups will also look at this and go, like, okay, like getting this actually into the hands of a hyperscaler is a lot more harder than just inventing the technology and showing that it works in the lab.

46:28 I think that is the key takeaway that you know uh your experience already shows.

46:33 Exactly.

46:34unknown Yeah.

46:35 Yeah. So what do you uh in the in the broader scheme of things? Like I want to step back a bit. Um we are in the era, I think, where we really don't have uh one solution fits all uh in the scale-up domain. I think we are at a very exciting time in history, in data centers and interconnect technology because all these years, copper has been the way, and everybody's like, yeah, just use copper.

47:01 It's a spine, and you know, in the future they're like, okay, let's put a PCB or something and hook it up, but it's all copper, however, you want to do it. Um we have come to a point where by the end of the decade, we will definitely not have copper working in the scale-up domain. And this is uh this is a first in a sense.

47:22 So I want to get your thoughts on what you think, what kind of solutions you think will emerge. And if you think one there would be one solution that fits everything, like this technology wins it all, or you think that there will be like separations of people trying various technologies, some uh hyperscheletors adopting solution A, solution B.

47:47 Like what how do you think of that as a competitor in the space?

47:50 I I I believe three technologies, our solution, optics and copper, I think they will coexist for long periods of time. Because copper is still useful for very short distance. I mean, some people they want to they want to use copper, so they will design their system to accommodate the short distance of copper.

48:14 And that way they can they can keep using copper. And some people they already heavily invested in the optical solution, and they must use it. So they will use it. And and some people who have kind of more sophisticated idea, I mean, they they will think of our solution as more reliable, cost-effective, and less power solution.

48:40 So they will use it. So I think for a long period of time, they will use three of solutions all together in different places.

48:47 Amazing. And do you see RF ETube technology also scaling to like faster um uh lane rates as we go ahead in the future?

48:55 Yeah, so we already are working on the next generation. So per cable core, it'll be close to one teraby per second data rate. So we're already working on it. So that'll be our next generation tips and eTube solution.

49:09 Fantastic. Like, yeah, that's that's always the question to ask because at some point we'll be like, oh, no, if if if eTube runs out of like steam um just as in uh copper, then what do we do? We have to then resort to optics. But the nice thing about wave guides, and this is more of a like an electromagnetics thing, is that there is no real physical upper limit to frequency in an opt in a waveguiding system.

49:36 Exactly, exactly. I mean, if you think about copper, why copper has limited bandwidth? It's simply because copper has the too many electrons. And those electrons, when it falls, it creates a skin effect. And because of skin effect, bandwidth is limited. And so what I'm saying is that the bandwidth limitation simply comes from the existence of electrons.

50:05 And this plastic, it doesn't have electrons. So in terms of bandwidth, it is unlimited. Bandwidth is unlimited. So once our silicone works, we don't have to really change the this cable at all. Just scale this silicon, then it will it will be used for higher data rate.

50:29 That's fantastic. Okay. That I think like this, your coverage of this topic has been very exciting. Uh, but uh just before the show, I wanted to hit one more thing. Just before the show, we were talking about uh some things like you know how we got into this technology and all that. And you mentioned that in spite of all your technological advancement on this uh you know interconnect technology using RF, your first customer was not related to RF at all.

50:58 Can you tell us that story?

51:00 Oh yeah, so I have been thinking about data center. That's why I started this company using eTube technology. But our first customer, they were uh display companies. I mean, not just one. So for example, they wanted to have the very thin wallpaper type of display. And you know to make a like a thickness is very thin, three millimeter, and they cannot put like a silicon in the panel, it's the OLED uh panel, and then they cannot put decoder, encoder inside the panel at all.

51:42 So they put every TB function outside of the panel, they put that in the setup box. So between setup box and the panel, they need high-speed interconnect because they cannot decompress and they cannot uh decode uh in the panel at all. So they need to transmit raw video data, uncompressed video data to the panel.

52:09 So data rate suddenly becomes 100 gigabit per second, 400 gigabit per second. Very high data rate for consumer electronics applications. So they look for some other interconnect other than copper. Copper is too thick, they cannot really fit in, and they look for optics, but optics is too expensive for consumer electronics and also it's power hungry.

52:34 And the uh in terms of thermal, it's too much uh thermal dissipation. So they don't want to use optics. So they look at our solution and they thought, oh, this is the one very thin, low power, longer distance. So they put some money for us to develop the TV application, consumer application, eTube solution.

53:00 So that was our first customer.

53:02 That's amazing. Is is that actually adopted, or did they find another solution out of it?

53:08 Unfortunately, they kind of uh delay their project and they uh stay with the uh kind of copper solution.

53:18 Yeah, yeah. No, that's that's a fantastic thing. So it you know, it gave rise to a technology that now has potential uses in something much, much larger than setup boxes and TVs. Like the data center interconnect is a far bigger opportunity than any of that stuff. So it's good in a way. It's okay if they didn't adopt it, you know, at the end of the day, you have a far larger opportunity because of that one problem you tried to solve in a different area.

53:47 Yeah, exactly. But I think I think eventually our generation, I mean, people will use more data in the future, not just data center. I mean, for example, one of our investors, uh Bosch. Bosch is the number one automotive uh kind of component company. And I think they're they're uh they invest in our company because we have a very cool, long-distance and low-power integrated solution.

54:17 I um they're thinking of using our YouTube solution in their automotive network. That's interesting. I mean, in in near future, I'm pretty sure we'll be in the automotive market as well. Of course, we need to spend a lot more cycles and spending time with our customers in data center first.

54:36 Yeah, that's a good reach application, right? Like a car cannot be 10 meters long. It's uh a little bit much, right? So the reach is limited in a vehicle, and that's what makes this uh good solution.

54:48 Right. So our cave is lighter and the distance is longer. And self-driving car, it requires many sensors and very uh uh high-resolution camera, and then they need to connect all those sensors all together to their ECU. So data is getting higher, and then they are looking for lighter because a lighter cable means a lot for automotive because it'll it will increase the efficiency of automotive.

55:17 Yeah, so this there's applications far beyond data centers as well. Like we started with obviously set up boxes, but now they could be automotive uh or anywhere you need like a high-speed interconnect uh at using relatively simple materials uh and technologies and whose reach is within the five to ten meter range, it fits.

55:36 This solution fits all of those applications, you know. Exactly. Yeah, awesome. Thank you so much, Sean. This was very interesting. Thank you for teaching us all of this cool technologies you're working on. Uh, definitely check out Pointu Technologies. Your website uh has all this information that people can go check out.

55:53 And you're you're often at conferences, so if any of the people listening here, they should stop by and check out your demos like I did at OCP APAC in Taiwan. It's really fun to see this stuff in action.

56:06 All right, Vic. That was a fascinating conversation. Now, what I'm super interested in is your perspective on will this actually replace something in scale up, some technology like passive cables or AECs or potentially rack-to-rack optics. And also, like, what do you think on the practical timeline of a new technology for scaling up like this actually getting deployed in volume?

56:34 Let's start with the second one first, because in terms of timeline, this technology is relatively new. It's it's old but new. You know, it's been around for a long time, but it's uh not been in deployment, right? And now uh you are seeing a lot more alternative technologies coming up. So we would put it at least, you know, a few years away.

56:56 Like I'm not exactly sure how to pinpoint it because it depends on how much they can scale to volume and things like that. But it's definitely not something that you will see mass production in hyperscalers or big data centers today, right? But that being said, uh today most of the scale up in a in a rack happens with copper.

57:15 And it's okay at 200 gigs per lane. Uh you can even run bidirectional, which means you can run transmit and receive on the same copper cable. It all fits within a single rack, no problem. We are now going from single rack scale-up domains to multi-rack scale-up domains. Now distances are increasing. And now when you put racks next to each other, now you need to hook them up, you need much more reach than copper can even provide.

57:40 Um, and now the question is, what are you going to do? So, are you going to use optics? Maybe, but it's very power-expensive if you're going to use pluggable transceivers. So the option is okay, let's go to co-packaged optics or near-packaged optics. Yes, but that's an entirely new ecosystem that nobody's ever done before.

57:56 So it's like a big step up from what we know how to do. All right, so you've got the copper doesn't work, optics works, but it's overkill perhaps. And so what are you left with? You're left with alternative technologies like uh micro LEDs or microvixels. Okay, now the problem with these technologies is that they don't run at the speed of copper, which means to say 200 gigs per lane runs within a single rack today, even bi-directionally.

58:23 These technologies don't do that. So you you know microvixels could probably run at 50 gig and micro LEDs run at maybe three gig, you know, so they are much slower. So what you have to do is you have to convert the 200 gigs per lane that's coming out of the XPU or the uh switch, and you need to convert that into something much, much slower.

58:43 And that is done with something called a gearbox chip. The gearbox converts is fast but narrow to slow but wide, and that has its penalty because this chip needs its own energy to make this conversion. It's another point of failure, potentially, and uh it I I'm not sure if it adds latency, but I guess the complexity is that it's just higher now, right?

59:06 So the beauty of RF over uh plastic waveguide is that it runs natively narrow but fast. So it's very much copper-like, right? It's not micro LED, it's not microvixel, it's not indium-phosphide lasers. It's very much copper. And if it's very much copper, uh it uses very simple electronics, like you know, we looked at it in the podcast.

59:32 Uh, it's just a chip and it just goes over a plastic cable, and that's about it, right? There's no complex integration, there's no putting three, five materials like lasers. You know, you can't make silicon produce light, and so you have to put all kinds of other materials, indium phosphide, gallium arsenide, gallium nitride on top of silicon.

59:49 So much complexity, so much integration problems. This does not have any of that problem. All you have is a silicon chip that's not even like cutting-edge silicon, it's like 55 nanometers. I mean, we could go to 22 nanometers and still get so much benefit, but it's still not even a FinFET, let alone you know a gate all around 3 nanometers.

1:00:08 No, it's none of that complexity. It's a simple chip, it works on simple electronics that is based on radio frequency stuff, which is very, very well established over many years of like chips and phones, and we know how to do this stuff, okay? We really do. And you can run this uh without any gearboxing chips because it runs natively at 200 gigs per lane.

1:00:28 So you can just like put it end-to-end, and if it has like, I don't know, uh up to five meter reach, yeah, you can at least connect two racks. And if you can have three meters, you can even connect up stuff within a single rack, and plastic is like much less weight, and uh much more uh uh density is possible within a single scale-up domain.

1:00:51 So it has all these advantages uh that like the optics world doesn't offer right now.

1:00:57 Yeah, interesting. Okay, so I like you that you make the point there's no 3-5 material, there's no indium phosphide uh bottleneck, this is just silicon, it's uh older technology that you could get built at uh foundational node, you know, um at many uh sort of trailing edge uh facilities. One question I have so is that um the antenna, is that also just silicon base then?

1:01:23 Yeah, it's completely integrated on the chip. So the beauty of this is that because of the frequency being 90 gigahertz or like 225 gigahertz, the wavelength of uh the the wavelengths are actually pretty small, which means that you don't have to have antennas external to the chip. Because antennas are always a function of wavelength, like a quarter wavelength is could be an antenna, right?

1:01:46 Uh half wavelength. So we've got different kinds of antennas and stuff. So you can make these antennas on the chip because they are actually small enough at uh the close to 100 gigahertz frequency. So everything is integrated, like you've got like transmit and receiver just antennas, and they're all on the chip.

1:02:03 So all you do is take a chip and put it into the OSFP module. Done. That's it.

1:02:08 Fascinating. What's very intriguing to me is like Credo sort of invented this new market with AECs, and it was pretty much just plug and play. You know, you just use the copper cable, but it has some you know DSPs to help make sure that that signal can transmit further and be cleaned up and retimed if necessary and so on and so forth.

1:02:33 But anyway, to the actual data center operator, it's sort of just plug and play, it's just like buy this and use it, everything works. It kind of feels like what you're saying is that not only is um these active RF cables um simple to make, and like conceptually, um there's no big concerns about the materials or bottlenecks or anything, but it also feels like it's kind of just plug and play.

1:02:56 Is is that how if you're a data center buyer, would you would is that a fair statement?

1:03:02 Yes, it's completely plug and play because the electrical interface is the only thing that the rack will see, right? As long as you can take the the plug goes into the electrical side and it works, what happens beyond that is not anybody's problem but the cable maker. Sure. Right? Your cable maker's responsibility is to take this signal electrically from one place and deliver it electrically in the other place.

1:03:28 Whether you want to do that with micro LED or like optics or like RF, it's not anybody's problem. The concern is is it reliable? I think Sean mentioned in the podcast, right? Like um, they are going to put this to the test, right? They want to make sure that it works. Because if you're gonna pitch anything to the scale-up domain, stuff has to work because there are way, way, way, way too many cables even in a single rack.

1:03:54 We are talking about 144 GPUs, we'll have above 20,000 cables. You can't have any one of them fail, and that's a single rack. Now multiply that amongst racks, right? Like, you know, how many racks in a data center, how many GPUs are interconnected? The reliability requirement of the cable is extraordinary.

1:04:13 And I don't think some this has been something that any technology has proven out uh at scale. Like, we don't know if CPO can do that, we don't know if NPO can do that. No technology has that data right now. Like copper is the only thing that has worked, and we know that we can deploy a data center full of copper cables and it's going to work.

1:04:36 So ultimately, data centers and like hyperscalers want to stay in copper as far as possible, but it's becoming extremely difficult. So the search for a new technology is on.

1:04:47 Yes, yes. Okay. Well, it sounds like reliability is the big thing. It sounds like the technology checks out the my thoughts around the cost being lower, that probably checks out just with like the simple um process node and keeping everything silicon. Um and even the plastic waveguide was quite simple.

1:05:05 Um, and and and then the mention of like it actually reducing the amount of connections that you need, cables you need. So it feels like the cost checks out. Really, the question is reliability. And I think to your point, we're early in NPO deployments and CPO deployments, even like rack to switch or whatever.

1:05:21 Um, but obviously the big question for point two and for this technology broadly is will it be reliable when there's you know hundreds of thousands of these deployed and running constantly?

1:05:32 Yeah, I fundamentally love the idea that there is another option that is not copper and not optics. And if it's RF, it's a bit closer to my heart because that's my whole expertise that I have built over a couple of decades in the semi-industry. So personally, I'm rooting for this to work just because it's RF and I love RF.

1:05:50 So uh let's see. Yeah, exciting stuff.

1:05:54 Exciting stuff, let's see. All right, listeners. I hope you enjoyed this conversation with point two and our thoughts afterward. Um, leave us comments below, send us an email, and we'll see you in the next episode.