The Real AI Infrastructure Bottleneck Isn't Chips, It's Electricity | Selva Murali

Russ Fordyce (00:01): Welcome back to the winner circle, everybody. I am here today with one of our all-star judges, Selva. He is a an engineer at AMD, one of those guys that is designing all the chips that go into our AI factories and all our computers and all these devices that we have. Selvo, welcome to the winner circle and thank you so much for being a judge. Selva Murali (00:19): Hey Russ, thanks much for the awesome intro and glad to be with you on this platform and looking forward to the chat. Russ Fordyce (00:25): Yeah, I I'm always I'm always so humbled by interviewing judges because you know it's it's you guys volunte I they're largely volunteer judges. we have volunteer judges and advisory boards and the the it it it just amazes me how awesome you guys are because there's like this is this is you got you're really at the heart of AI. You're building those chips that are going into the data centers that are powering AI before we get into all the nerdy stuff. Tell us a little bit about you and and what you were all about self. Selva Murali (00:58): Sure. And before that, thanks for putting together a cool pat platform that as judges and you connect with a ton of companies and rank them and deliver them awards for what awesome work they are doing. That's really a phenomenal stuff that you're carrying out. Yeah, from what I do, I am currently a staff level engineer at AMD, wherein I work on some of the leading edge technology nodes like two nanometer, three nanometer or even beyond that like one nanometer and stuff, mainly for our data center scale products and for every other pretty much anything and everything that the company delivers in terms of silicon. I also am working on a ton of projects that involves adoption of A and ML for circuit design and also for more interesting semiconducted design to make sure we squeeze out every small bit of power, performance and area. That's mostly that I do and I was fortunate enough to be able to file a couple of patents and also have a published almost over like ten to eleven technical papers currently, which does have some ninety citations. that's what my Google scholars say, so I count on that. So that's where I stand and I'm looking forward to staying in this industry for quite a long period. Russ Fordyce (02:23): Well, you you're you're you're dead in the right place, right? I mean this is this is where all the money is really being made and all the all the i yeah, well it's I mean it's it's where the digital meets physical, right? And that's what I love about this. I was always I was in IP for, you know, in you know the internet protocol and Ethernet all my life, all my career. And at the very end I got into light and lasers and physical and chip. Selva Murali (02:30): Yep, I hope so. Selva Murali (02:36): Mm-hmm. Russ Fordyce (02:50): And it was very different because we weren't selling services and you know, we were now selling physical things that moved electrons. And that just fascinated me. What drew you into kind of physical design and chimp implementation specifically? Selva Murali (03:05): Uh-huh. So first thing, so there are mainly two phases in the silicon design. One is the front-end design and the other is a back-end design. So front-end design is a space where people write a lot of they kind of decide what's the architecture, whereas a back-end team takes the architecture and implements them into circuits. Ever since childhood, I always was a circuit-driven person, but not the architecture-driven person. So that eventually Pushed me towards doing a master's in the circuit phase and learned a ton more skill set during my masters, wherein I got a wider picture of how this industry works and how the front end and the back end team interacts. And I felt like my sweet spot would be somewhere in the back end team, but not purely into the back end, but somewhere in the intersection where the front-end team ants over to the back end and they take over. So that's the spot I wanted to be. And luckily I was fortunate enough to be able to secure an opportunity that lets me do what I wanted to do. And eventually that's what I'm doing for the last five, six years. Russ Fordyce (04:14): Well, it's it's it's the it's really where all the constraints are coming. If you think about AI and all of our challenges with AI and data centers, both good and bad, they're all right down to the chip because you like your role, I bet your only goal is to make the chips faster and more power efficient. Ever every year, right? It's got Selva Murali (04:18): Yeah. Selva Murali (04:25): Mm-hmm. Selva Murali (04:34): Yes, yes, that's what yes, that's mostly I focus on. Russ Fordyce (04:38): Yeah, and it's it's such a battle. It's it's literally, I was so amazed that we, you know, we had a new marketing term which was cost per bit. Selva Murali (04:47): Yeah. Cost and yeah, basically TCOs getting called as cost per bit. Yeah. Russ Fordyce (04:52): It's it's like the lowest lowest metric you can get to is cost per bit. How do you it that's all n good kind of fun nerd talk? How do you explain what you do to somebody outside of some semiconductors if you're at a dinner party with friends or something? How do you explain it? Selva Murali (04:55): Yeah. Selva Murali (05:08): So, in simple terms, I'll just say like I work on making sure chips are faster and your laptops have a longer battery life, and you don't always pay a ton more cash for upgrades. Basically, I make sure your chips are smaller. So we pay smaller money to the foundries and eventually you pay smaller money to us to get a more upgraded version. So that that would be my simple pitch. Russ Fordyce (05:34): Yeah, and and people are always surprised. I was surprised that just how prolific chips have become. You know, there's we joke about like you've got a chip in your toothbrush now, you know, you got an electric toothbrush. but it it has, and I I was talking with another semiconductor manufacturer and Selva Murali (05:41): Yeah. Yeah. It it's gone crazy, like everything we touched yes, yeah. Russ Fordyce (05:55): I said there's gotta be, you know, a hundred or more sent you know chips in a car now. And he said, there's probably hundreds and we looked it up and it's you know, it's thousands. There's five or six thousand chips. Selva Murali (06:03): Yeah. Yeah. Automobile is like fully controlled by chips now. It's like it's no longer a mechanical engineering driven industry. It's more of a mechanical plus electronics driven industry. Russ Fordyce (06:16): Yeah, the I think the luxury cars had, you know, five to six thousand chips in them, which is just amazing. Selva Murali (06:19): Yeah. The luxury cars are like purely not even like no longer mechanical engineering driven at all. They are like purely into the ADAS and which is purely controlled by semiconductor. Russ Fordyce (06:31): Now you've moved between Samsung AMD. They're kind of two direct competitors. They kind of approach different markets sometimes. I'm sure they approach chip design different. Is there anything that you kind of have noticed about or what have you taken that's good out of both companies in terms of kind of philosophy of chip design? Selva Murali (06:37): With me. Selva Murali (06:47): So both the places have a pretty good engineering talent. And that's something has impacted me positively. And I'm happy to be in these enterprise great companies where you get the exposure of cutting edge technologies and at the same time have the opportunity to work with some of the industry leaders. So, in terms of what's the key difference, I would say like Samsung. touches pretty much everything, like all the way from memories, from manufacturing their own chips using Samsung Foundry and designing their own chips using Samsung electronics and making sure it goes into their mobile phones and eventually make a mobile phone and sell it to mobile phones. Also make data centers, they touch everything in the ecosystem. Whereas AMD, they don't touch everything, but they want to get a specialization. Where they focus on just the data centers and the client scale devices, which is basically your laptops and whatnot. So that's a key difference. It kind of required me to have a change of mindset and how I looked at problems because in Samsung it's more of a holistic approach where you have everything under one roof and you have a little more control. With control, you get to have more challenges in terms of convincing people. Whereas because no longer you are working as a vendor and a customer relationship, it becomes more like teams. You are working with teams and everyone has their own preferences and you can't any longer say, like, hey, we want this, get this done. But whereas in my AMD experience, it has been more like we want this, if you get this, we'll buy from you. That was experience. So that's quite a different mindset approach that I had to make. But eventually I think It's pretty good experience that I was able to work with all under one roof versus things getting pulled in from different vendors and how to work with vendors and it's a holistic approach in my view. Russ Fordyce (08:52): Yeah. It's it is it's interesting that you say that a kind of that vendor versus internal team approach. I never thought of it that way, but as a you know, as a as a customer, you get to go in and kind of say, Hey, this is what I need. And yeah, when you're talking to a team member, it's a lot of convincing. Hey, this is what we should do. Selva Murali (09:04): Yeah. Yeah. And you also put some pressure in your Yeah. Yep. totally agree. And with the cus with the vendor you have the privilege of saying and the vendor is also working may to make sure like I don't lose this customer. Whereas if it's all internal, what happens is they know that you can't go out and buy something because obviously you need to pick something which is available internally. So they kind of have that leverage and say like, eventually these guys have to come to me, they can't go anywhere. So Russ Fordyce (09:40): Well, now you say that. And what's funny, they did buy it, but I was in telecom for a lot of years and we actually ended up selling ATT retail, the retail division of ATT. We sold them ATT wholesale cheaper than they could buy internally. So they would buy from us because we got a cheaper rate in tur from the wholesale team than they could get internally. So they bought from us, which was just ridiculous. Selva Murali (10:07): Mm. Russ Fordyce (10:10): You're really again at the heart of this kind of AI world. When people talk about kind of AI infrastructure bottlenecks, we see it on the news. We're talking about, you know, data setters getting shut down because nobody wants them in their backyard or whatever. And we hear about NVIDIA producing gazillion GPUs and AMD producing GPUs and everything like that. What's the real bottleneck that kind of nobody's talking about? Selva Murali (10:22): Mm-hmm. Selva Murali (10:34): So something that I sense and I'm reading about and what I work based on what I work on, I strongly sense like power would be the biggest button like that we'll be chasing upon. And that's something we already sense in the structure how these data centers are being positioned. They go to cities where they can build their own power grid and give their power grid exclusive to the data centers. So that clearly sa shows us like power is going to be the biggest bottleneck as these comp as these data center companies keep building, like, as you said, gazillions of GPUs put under the roof. They all are power hungry and they consume like hundreds of thousands of apartments worth power every day. So power would be the biggest restriction or biggest limitation for scaling. And second thing obviously would be managing the heat because Managing the temperature that these data centers are gonna generate. It's like even with our own mobile phones, we can sense like if it's getting too much overutilized, you can just draw parallels over here. You can sense like your mobile phones get too odd and you can't even keep it in your pan pockets. Imagine like like billion mobile devices getting heated up at the same time, the entire temperature. Russ Fordyce (11:55): In the same room. Selva Murali (11:57): Yeah, in the same room as well. So that's where your cooling systems w takes up a biggest market share in these data center success. Russ Fordyce (12:08): In in this chip design world, what happens when kind of your physical design doesn't hit its power or performance or area target? What's the kind of real world consequence down the line? Selva Murali (12:20): So first thing it starts with like you're gonna be slower in terms of performance. So what happens is as you become slower and slower, you need more GPUs to be added. Because earlier you projected like with 10 GPUs, I can deliver this much terabytes per second or flops per like flops per bit I am able to deliver. But as you are struggling to meet the performance, what happens is Earlier you were projecting 10 GPUs, now you ni you require 12 GPUs. So the cost increases for the customers, and at the same time, the number of GPUs that you need to plug in, power them up is gonna increase. So basically, your power profile increases. The total power that you budgeted initially, let's say you budgeted one kilowatt hour. Currently you're looking at like 1.5 kilowatt hour to get the same flops per bid. So those are all the real world consequences. And eventually you'll be spending more on power, more on area to plug in two more GPUs. So it all adds up, everything comes down to the financial part, but indirectly you influence the environmental sector as well because of the all the heat generated, power consumed, everything connects to the total ecosystem gets quite disturbed as we fail to do a good job. So that's where we are always kept on the toes to make sure you do a good job. Russ Fordyce (13:47): Why is it why is optimizing the kind of physical layout so much harder today on a chip than it was ten years ago? Selva Murali (13:53): So it starts with the earlier 10 years ago, we were looking at like something like 48 nanometer, and and at 48 nanometer you didn't have these challenges. And also semiconductors, the amount of chips that were consumed by the global market was very small. Like 10 years before, I think the total market capture was like not even like 50 billion dollars. Like currently we are looking at a trillion dollar market, which shows like the Total amount of chips that people consume directly and indirectly. As like for instance, you were saying about ATT. ATT people may think like I'm not buying any chips, I'm just buying the mobile service from them. But for them to give you mobile service, they consume a ton of mobile processor chips. So the semiconductors that people consume has grown up crazily, and which in turn led to a ton of innovation where we are currently down to like one nanometer and two nanometer semiconductor devices. I think in the future, in October W W D C of Apple, I'm sure they're gonna come and say like the next iPhone is going to be made in a two nanometer technology, which shows Russ Fordyce (15:03): Right. Yeah. That's a and that yeah, no, but that's a half of a millimeter, right? So you you think of a centimeter, think of a tenth of that and a half a tenth of that. And now we're talking about going to two nanometers. Selva Murali (15:11): Two nanometer. Yeah. Selva Murali (15:16): Yeah. Mm-hmm. Yep. Yeah, two random meters like we can't even see it visually. Like it's like it's beyond microscopic scale. So the problem that arises as we scale down is you need to be more vigilant of the rules that you need to follow. Because if I go and give them a random layout, the foundry will give me back a random silicon, which will not work eventually as we intended. So we need to be extremely cautious in following the rules that the foundry Russ Fordyce (15:28): No. Russ Fordyce (15:32): Yeah. Selva Murali (15:52): When I say foundry, I'm referring to like TSMC or like Samsung Foundry or Intel Foundry or any other foundry. We need to be more conscious about how and where they are structuring and limiting us and follow those rules rigorously such that you don't get back a scrappy silicon which directly goes to trash. So that's where all these challenges come up and scaling down is has both its pros and its cons. So we look at the cons. Russ Fordyce (16:12): Yeah. Selva Murali (16:21): And making sure the pros are actually consumed rather than being disposed. Russ Fordyce (16:27): Now you mentioned your some of your research. One of your research reports looks at kind of how machine learning can be used to automate kind of that physical design. Walk me through how you kind of validate that that ML driven approach is better than the human one. Selva Murali (16:44): Mm-hmm. So everyone is adopting ML techniques, and something in that line of work is what I was trying to do, and it kind of panned out well. So basically, the outcome of that ML-driven work was let's say a human is trying to do physical design work, wherein he's trying to place 10 AND gates or R gates, like basically your R logic and AND logic gates. If he's trying to place 10 gates in a given small area, he has a limitation of the number of permutations and combinations that he can make. And obviously, he gets at some point he gets frustrated, hey, this is the best, and he packages it out. Says like, hey, I have tried my 10 times, and this is the best I can place these 10 R gates together, and I this is the best performance that I'm able to deliver. But with these ML treatment techniques, what happens is I tell the tool, you do any number of iterations. I need the best timing, which translates to performance. As you improve timing, you get a better performance. Basically, you're able to clock your designs more aggressively, you get a better performance out of it. So I tell the tool you take n number of iterations, I tell like 1000 or 2000. The tool goes and does like 2000 times of these 10 gates placement, and it gives me out of these 2000, hey, I find this to be the best one. So that kind of removes the human psychological barrier firstly because it's not that the ma human individual physical individual cannot do that he can do that he or she can do that but it's gonna take them a long period of time probably like six months whereas this ML-driven approaches do does it like in a day or so so that's the biggest difference and basically you leverage the compute to do more variations and more optimizations and get back the best performance and if you are able to let's say earlier I told the tool use 10 gates. Currently it comes and tells me you don't even need to use 10 logic gates. You can just use eight and you'll be able to get the same outcome what you are intending. So those are all some of the findings that was eventually drawn out of those technologies that I was working on. Russ Fordyce (19:02): What surprised you the most when you started applying all those techniques to actual, you know, that physical process? Selva Murali (19:08): Mm-hmm. First thing the setup was complicated. So, in order to make sure it actually works, it took me a ton of time. Probably like I I would have invested almost a month to make sure it works. But once it started to work, I kind of scaled down my manual inputs to the tool. Like earlier, I was doing like let's say like two days worth of work. It boiled down to like 10 minutes down of the work. So it kind of saved me a ton of time. Probably I'm That's the reason I'm able to do a ton of coffee chats with my friends, like speak with you, for instance, and like hang out as I like until my superiors come and say, like, take up a new task. I'm able to spend my time more productively, I would say. Russ Fordyce (19:55): Yeah, well it's what that's what it's all about. It's I I don't it's this is not about losing jobs. It's about making us all better at what we're already good at doing. Yeah. Or or ext or extending our capabilities. We were talking about that before we started recording. You know, I'm I'm now a developer all of a sudden, you know. It's it's it's really different. And you know, it's I've been using, you know, Claude a lot. And, you know, there's a design tool inside of Claude now. I'm a graphic designer by trait, you know, by Selva Murali (20:00): Yeah. Yeah, leveraging. yeah. Selva Murali (20:11): No. Selva Murali (20:24): I see, I see. Russ Fordyce (20:24): by training originally as well. And boy, that's, you know, now I have more design capabilities and prototype capabilities. It's it's just, you know, it's a I'm a different person now with with AI. You're you're facing a different endeavor though. And what you said is really interesting because you're what would have taken you six months to build now can be built in a day or a week by the AI, but you're still having to put in a week's or or you know a month's worth worth of instructions. Selva Murali (20:31): No é top. Mm-hmm. Russ Fordyce (20:52): Because you would have had a month worth of instruction if you had done it yourself, right? Of that six months it would have taken you, it would have been a month of figuring all that stuff out. And that still has to be done, which is really interesting. But you've done that. It's a and I'm gonna I'm gonna go off on tangent real quick. It's like borrowing somebody's spreadsheet. You know, when somebody's got a really complex spreadsheet, you gotta try to figure out how it's all connected. It's you're like, huh, you're doing the same thing with this ML. You've now got to go to your design team. Selva Murali (21:01): Yeah. Yeah. Selva Murali (21:16): yeah. Yeah. Russ Fordyce (21:22): And go, hey, I built this ML to to do this better. and it really is great. How do you build trust within your design team to make sure that they understand that this process is Selva Murali (21:32): Yep, that's where yeah, yeah, that's where data is our best friend. And as we always say, that's the reason I as I was telling, like I try to build a ton of data by doing these trials and like runs over, runs over runs and show them like see our standalone, like your conventional methodology was able to deliver this much. My ML-driven approach can deliver the same, but at the same time it can also deliver something better if you give it like Some compute, throw it some more compute and ask it to do like hundred more iterations, it can give you a better outcome in a lesser time than what your conventional methodology does. And most of the times, at least in the places that I have worked with and the leaders that I have worked with, they were all pretty much receptive towards change and they have been willing to adopt change. In most places, bringing in change and making sure the people Adopt that change is the biggest challenge. But I was fortunate enough that I need not handle that type of red tape world to actually propagate. But as long as I'm able to demonstrate that this is true and this is realizable and data goes along with it, people are willing to buy it as we pitch them. Russ Fordyce (22:52): Yeah. Now in your world, the that kind of there's always that famous kind of trifecta of Venn diagrams of power performance and then kind of the just physical area of chip design. It's kind of always the trade-off between those th two and you're looking for that dead center to hit the project. It's hard to optimize all three at once. It's like good, fast, cheap, right? You can do two. It's hard to do three. When you're when you're thinking through a project, how do you think through that process of power performance or area? Selva Murali (23:04): Yeah. Selva Murali (23:10): Mm. Selva Murali (23:23): So it depends on the end application. For instance, let's say if it's going to be CPU, CPUs always focus on performance. Like you are willing to spend any amount of area, the first priority would be performance, and then comes your power. And area is like pretty much if you're able to save well and good, but we don't care about area. But whereas when it comes to GPUs, GPUs, the way we approach GPUs is like, You can plug in more GPUs if needed, but we can't give the power that the GPUs are consuming currently in the rate that we are scaling the GPU infrastructure. So when it comes to GPU, all that we focus primarily is power, then comes area, and then comes performance. Because these GPUs are large, like they have become like so big that you need to actually start to limit the area that you throw at building these GPUs. So that's where we Prioritize power, then comes area, and at last comes the performance for the GPUs. Whereas for CPUs, it's like performance, power, and then comes area. So it's a different world for different applications, I would say. Mm-hmm. Russ Fordyce (24:31): Yeah. Have you ever made a kind of a design decision that you kind of looked at and were like, maybe that's not the right way to go and it turned out to be the right way? Selva Murali (24:40): not until now, probably in the future. I'm hoping it doesn't happen. Yeah. Russ Fordyce (24:44): I've made a couple of those where you're like, maybe I shouldn't have done it that way. And then you're like, I'm glad I did. I it was good. Now AMD, you're you're there. Yeah. Now you're you're at AMD. They're just launched that MI 400 series to go go right after the NVIDIA stuff. You sit inside of this kind of chimp implementation. What does it actually take to get a chip like that from kind of design to something shipping? Selva Murali (24:53): Yeah. Those are good learnings, I would say, yeah. Mm-hmm. Selva Murali (25:13): So first it's a very long process. So all these activities go like take like at least two to three years. So if you're seeing some product being launched today in the market, it just means like these people have been working on it like three years before, writing the specifications, architecture, and then they come up with their first front end implementation, and it takes a long process to make sure the design is realize and then switching them to circuits, then sending it to foundry. It's a very long process. So it's like it takes a long period of time firstly and requires a ton of investment. Like we are looking in the ranges of like hundreds of millions of dollars goes behind making sure the product is firstly they figure out if this product is worthy of investing hundreds of millions of dollars. If it is they also look at the total addressable market, the TAM. That these products would be able to generate for the company. And if it both balances out, we go full in on, invest all the resources, be it human resources or like the compute resources, tools, licensing, everything, and make sure it's the best in class product that we're able to deliver. And that's how you look at all the products. So it's a long period of time, and ton of engineers spending their days and nights during weekends they do work and everyone has worked during the weekends for multiple of these projects that I have been closely involved with. So that's all eventually paying off in terms of better products being able to be solved and of course better stock price and better compensation, I hope so. Russ Fordyce (27:00): Yeah. Fat faster everything. I I remember from my semiconductor days, I think we were looking at like three years from basically like that initial kind of design all the way through to production and a shipping product. And that's where we're getting stuck a lot now in the semiconductor business is not in the chip fabrication, but it's actually in that packaging of getting those chips into boards and into into equipment. Selva Murali (27:11): Mm-hmm. Selva Murali (27:23): Yep. Now we have added more spaces. Mm-hmm. Russ Fordyce (27:28): Yeah, we the the company I used to work for just received a big grant to to basically triple their capacity for just that thing. Not not for the fabrication. I think they got some money for that too, but just for packaging because there's this huge bottleneck of those chips have to that the you know, they're just not, you know, pushed out on a on a chip and then sent out. It's the you know, they go onto boards and they go onto other things and into equipment. And that's really where a lot of the rubber meets the road is is really getting it into things. Selva Murali (27:39): Just for packaging. Selva Murali (27:59): Yep. Russ Fordyce (27:59): And getting the people and getting skilled labor to put it in there. That was one of the hardest things was having the skilled labor just to get those chips soldered and, you know, perfectly aligned and everything else onto those boards. It's a very interesting time. now you've been Selva Murali (28:15): Yep, slowly people are moving from software to hardware and circuits. I mean the youth are having some interest towards circuit side as well. Let's see how it shapes up. Russ Fordyce (28:26): Yeah. Well, I I yeah, I've I have adult children and you know I've been trying to push one of them into kind of that into that field. He was an IT guy and I'm like, good, you got that background, but let's go get some physical stuff too. Why don't you, you know, become an electrician or you know, something like that where you get kind of that hands on because that's gonna be very, very important. You you see a lot of these nominations, you judge innovation for a living. Selva Murali (28:44): Mm-hmm. Selva Murali (28:49): Yeah. Russ Fordyce (28:52): You you're at the center of it. You've you've judged cybersecurity for us and AI and customer service award. You've done kind of everything. What separates kind of a generally innovative submission from one that's just well marketed? Selva Murali (29:06): So first thing, many a times it's not always true that big company, if they make a submission, it's always like the best submission. I think some people get carried away, like let's say Cisco makes a submission. People think they Cisco submission. I'm just throwing a random name. Let's forget about Cisco for the sake of yeah. Yeah, ABC company. Let's say it's a Fortune 10. They think they tend to think like it's a Fortune 10, so it they need to be the best one and they tend to give the best scores. Russ Fordyce (29:22): Right. A B C Co. Selva Murali (29:35): But for me it has always been reading through the description and visiting their websites and seeing what their websites are actually projecting and what their product videos actually show. That's how I kind of make my judgment over here. And I think that has been pretty good and working out for me. Russ Fordyce (29:55): Yeah. It's well, it's it's it's such an interesting time because all of a lot of these sectors are actually combining into one. You've seen customer service take on a massive role in AI as well as sales and marketing. We have the SAMI Awards, that's now very heavily AI driven. You're in this semiconductor market. I think IDC puts semiconductor growth at twenty five percent in twenty twenty six, which like if you're an investor, that's a good thing to be in, right? Selva Murali (30:14): Yeah, yeah. Selva Murali (30:23): Yep, yep. Russ Fordyce (30:24): What does that kind of growth actually feel like inside of that engineering team when you're being, you know, kind of pushed to do more and more and more? Do you get more resources too? Selva Murali (30:32): Yep. So it's it's a it's a double edged sword, I would say. So people tend to s tell you that every day and and out you there is some chat around your stock price going on because most of these engineers get compensated using their stocks as well. It's not just your base pay salary that they get. Most of these engineers get something like a lump sum of stock units. So there is always a chat going around like Today I saw it as increased like 10%, 20%. Today it has dipped by 20%. Yeah. At the same time, from the leadership, there's another message coming in. You see it has increased by 20%. If we need to make sure it's consistent, you guys need to deliver. So that's the kind of pressure that you get passively because of all this growth. And for obvious reasons, we want to keep up with that growth and The engineers they themselves get motivated to squeeze in extra hours and like sit in the late nights and keep throwing in whatever the best that they can throw in. Russ Fordyce (31:40): Yeah. Now chip design's one of the most specialized, kind of longest training cycle disciplines in tech. How do you think the field should be attracting and training new engineers right now? It's this is a it is not a it is a growing field, but there's not a lot of people going into it. Selva Murali (31:57): Yep, yep. That's a that's a very interesting question, Russ. And especially with respect to the semiconductor industry, it takes a long period of time for the individuals to be trained, shown what you need to do, and most of the knowledge is ancestral, I would say, where you get it from your managers, your mentors, and you apply it. Whereas for software it has been very self driven. So if you can learn it by yourself, you can just learn and you can do and you can deliver products. Within like three months, you can ship a product and even you can make some revenue out of that product. So this is a long cycle. So people tend to be more hesitant towards getting into this field for obvious reasons. It's not like I'm complaining, but slowly this field has taken some traction in this decade, probably until the one more decade, I would expect. It's a cyclic world as we all know. So I Personally, I think semiconductor space would always be existing and growing, but not sure if it will be growing at the same rate that we saw over the last five years. Probably a slightly slower pace for the next decade, but it will always be growing. And one thing that I enjoy being in this industry is you tend to have the ability to influence a lot of decisions and at the same time able to push a ton of your own. interesting works that you might have done, you can really push it. Whereas not that's not always true in the software world, whereas there'll be 10 other people working on similar features and one gets picked and the other nine gets kicked out. That's not the case with the semi-gun space. And there is a ton of growth possible as you stay in the industry for the longer run. Whereas in the software space it has been more like extremely competitive space and growth is Crazy in the initial years. And as you scale up, it becomes very cutthroat in my view, but I could be totally wrong. So that's something I would always want to sell to my juniors and my the schools that I speak with. I always push them like try to explore the space. And if it's not for you, you can always go back to software. It's not like I write software every day in and out, and I should be always equipped to be writing software. Selva Murali (34:21): The the difference is I write software to design circuits and you can write the same software for designing some programs and algorithms as well. Russ Fordyce (34:30): Right. Well, it's I think it's a I I think your point's really well taken, which is there's there is there's been such a proliferation of software developers that it is very competitive and there's not a whole lot of differentiation. The ASIC design in this design is really thoughtful work, not that software design isn't, but it's very thoughtful work and it is it's a lot of it is very human driven right now. Selva Murali (34:44): Mm-hmm. Selva Murali (34:51): Mm-hmm. Selva Murali (34:56): Yep, yep. Russ Fordyce (34:57): and you really do have a chance to kind of be a superstar and really to put your mark on something. there's still a lot of innovation that can happen there. I don't think there's a ton of innovation going on in the software world, other than new types of products. You know, so it's I think I think that's very true. If you were if you were a college student going in, it'd be a very good place to kind of stake your career. I'll I'll challenge you on the declining number of circuits, though. I the I, you know, the people said that about internet capacity too. They said, well, we'll reach a point where we'll have enough. And you know, we can't get enough. We can't we can't we'll never have enough. So I don't think we'll ever have enough chips either. I think that ch unless we get to quantum real quick. And then that's that's a whole new new new exercise. Is there any misconception that you see college students making about kind of that career path into ASIC design and implementation? Selva Murali (35:27): Yeah. It we have never reached that point, yeah. Selva Murali (35:40): Mm-hmm. Yeah, yeah. Selva Murali (35:51): Mm-hmm. I see currently I even spoke with Arizona State University recently, and I also spoke at MIT recently, like I think two weeks before. I see a ton of N graduates more interested towards circuit design, and they're all getting to know what this industry has to offer and what they can offer to the industry. So I think the youth are pretty much motivated and they're no longer kept on the blind side that. this is not a great place to be in or like you have limited growth, you have limited compensation ability. They are they tend to explore more through all the open source data that's available. And I think I'm sensing a good amount of traction already and it's a ripple effect and it will keep on growing on its own is what I see. I don't think it's gonna drag anytime soon. Russ Fordyce (36:46): No. Now you've like I mentioned earlier, you judge kind of outside of your specialty area, which is great because a lot of these things are now kind of blending together. specifically around customer service, has did when you judge those, did it change the way you think about your own field at all? Selva Murali (36:52): You hear? Selva Murali (37:06): yeah, for sure. That's that's a very interesting space. So that I even for instance that judging kind of opened up a different view to me. Like how I used to think like customer service is just all about speaking with your customers and solving their problems. But with all this new scale technologies and new scale products, looks like it's a very different world and the way that you approach customer service has transitioned like drastically over the last five or ten years, I would assume. That's a that's a different eye opener for me and a big learning for me personally. Russ Fordyce (37:43): Yeah. What's is there something that you've judged that's kind of changed your approach to work? Selva Murali (37:49): I would say towards I would say yes, something. Yeah, yeah. I just recall like I'd say like most of these companies have transitioned to AI-driven customer service, and that's kind of paying them mostly, not for everyone, every one of their customers, I'd say. So most of their customers seem to have been paid off well. So that's something I'm seeing in my workplace as well, where we are adopting AI for. internal customer service driven work. Russ Fordyce (38:59): So you've judged not only kind of the technical and the techie stuff for us, but you also judge things like customer service awards. Did anything pop out there that's kind of changed your work? Selva Murali (39:11): Yes, earlier I used to view like my internal customers as more like non-competing or like I can offer whatever and they can they need to consume it. But later with all these approaches and I also have a full view of what goes behind the customer service. And I kind of realize I need to be a little more provocative in answering them rather than treating them like They need to come to me and I can offer them this fix anytime. That's kind of as a as a mindset in change after these judging activities. Russ Fordyce (39:53): Well, it's funny when when I was I've worked for a few companies where they actually put you into the field, even as a marketing or you know, kind of working at a corporate headquarters type of environment, they put us out in the field to make sure we understood the business. I remember hearing about a friend who went to work for UPS and he was, you know, some sort of executive trainee, but they put him in a truck for six months before he even got to go into, you know, put a suit on. and I always thought that was really good because you got that practical experience of seeing what you know what the customer experienced. And I think you'd you see that too, even, you know, maybe you're not on the AMD customer service lines, but you see the amount of technology going into that into those into those those jobs, into that sector. Yeah. And the AI impact on that sector is just enormous. It's just it's crazy. Selva Murali (40:34): Yes, yes, yes. Yep, totally agree with the press. Selva Murali (40:45): Yeah, yeah, yeah. Russ Fordyce (40:48): Well Selva, thank you so much for joining us and thank you so much for joining all the you know, for judging all these nominations and being part of the part of our community. Selva Murali (40:49): Yep, I totally buy that point. Selva Murali (40:56): Yep, of course. thanks for all that you do, Russ. Russ Fordyce (41:00): I appreciate the time. Thanks again. Selva Murali (41:03): Mm-hmm. Thanks, Russ.