Yeah, So the other thing I want to talk a little bit about is some of this stuff that's happening in the electronics world that I thought you might need. There's might find interesting, but obviously you all know about Moore's law. It does have a very very very same exchange. It can be here.

Yep for for all that stuff. but Moore's law in terms of its traditional sense, actually it's not really going on anymore I Know it almost ended as it ended. Do you think here is it all my it has one run, almost a generation Yeah. I Feel yeah I mean inside already skipped one generation, rather scaling right? So more line is most traditional sense.

It refers to the fact that yeah I can simply double the computation of opacity or something because I can cram more stuff in it at the same space that it was before. But we're hitting those limits. Yeah, so once you need those limits, then obviously we can't just cram more transistors and expect things to become better. So then you have to change so now.

but there is still a more lot going on, but at a layer above. So now the computational capability of systems will continue to grow but not a dry but not simply riding the the transistor scaling lot. So in actual so how does that happen? So it happens by it not becomes about a hybrid of the system, right? So throwing our software a combination of software definition of hardware together and design techniques rather than having an access to the next technology is going to create that any technically gene techniques as well sleep. Yeah, oh my God.

did they double the memory on the back of a car stereo? No, Exactly. So ignited. Now your 3d integration. Yeah and so on.

And more and more process than you can even think about some of the paralyzation of Surahs forward integration because now you're paralyzed in time and anyway, things like that. So but yeah, that's kind of where you just headed I mean I don't design microprocessors. My work is on Millimeter-wave a 6-4 you know Wireless and optical systems. But so even for in my field this Moore's law now refers to the entire capacity of the system.

No oh my gosh. I related so to understood that is a ploy to yeah I mean you have to if you want in you because all you care about is computational capability of computational capability continues to go up and you can say that okay we're right in that yeah nonetheless. yeah. so it's interesting to see that this is still going to continue.

So if anyone ever tells you okay, Moore's law is dead. Yeah, okay, it is in one in one sense. Yes, then there's other thing is that the cost per transistor actually has even going down. And you know if that's right, If you want to do a new fabrication, run on a 70 nanometer bit of the memory is not going to get any cheaper.

No the cost of that except when they're a mitosis. Yeah, the cost. And yeah, so it depends on scale numbers, right? There are numbers if you want to make something only ten thousand or something and you want to put it in the state of the earthward it's going to be. It's just the first time you submit for a fabrication.

Looking at a couple million, right? So if you're not yelling and selling them by millions is just not the massed cost of production. It's just so high that in terms of Mikado's yeah, don't be question ever left field, the RF Hill will meet a wave stop yeah that you're working on. Yeah, is there an bad analogy, but is there like an FPGA equivalent in that sort of field? So so so they're different. Everything have to be totally custom from scratch or can you sort of depends on what you're lowering how much you're doing.

So so if you want to think I'm doing the analogy of the FPGA like you were saying. so this whole software-defined radio we have this very that is huge, right? The whole yeah, you probably know about software-defined radio from the little dongle. you can borrow things like that. but in reality, what started the fire radio in it in a real communication system is is that I can have one box that can work across a wide range of frequencies, kind of interface with a wide range of modulation formats and communication formats and interpret them.

So in theory, you could build a system that behaves like an FPGA as a hybrid row. Not a single chip. Has anyone build the software-defined radio? No, that's not. Maybe some in backhoe radio, they just really broadband the radios that can be, but they're not really quite the way you one can just program it and you know, do it.

I Mean write more custom-made yes, at lower frequencies, in the I assembled yet a quad band and whatever all those things you can do. But once you go into the higher and higher like at 5g, look at the next generation of communication. We're actually going to jump to 28 Pico, Hertz and 39 diggers and that'll be consumer level. Yes, now it's going to be the right Everybody and their uncle is working on factory radio naturally.

Okay, it's going to happen at some point. Is that because that's a more ideal frequency atmospheric wise? Or is it a regulatory been available? Hell that. I regularly that it's rates abandon allocated. So I think the 800 megahertz around 28 over 20.

90% of I/o in every country I Think some of them are actually quite international I'm not sure about in Australia but even highest frequencies in Japan for example, above 100 meters to the Bell right? If you want to use, the talking's become a lot more difficult. Yeah. I carry that I I Heard. In fact, atmosphere occurs that chip level difficult both.

Oh yeah I mean it's difficult for a couple of reasons. For example, 28 Gigahertz just doesn't go through your house as Riley if it's raining. the attenuation if it's raining is huge. and they also want to do phased arrays.

They want to do beamforming in the air. Yeah, so that they target individuals with a beam. and that is a totally different style of communication than broadcast modes that we are used to. Internet for second, so it's going to be completely different.

And yeah, so it's all I'd like to see how it's going to play out, but everybody is working on it and obviously Nokia is working on which is where I work in the labs. but uh yeah, everyone knows about it. It's not a secret, but it's going to be quite difficult and interested to see it. It's gonna have to be very inexpensive in order to become competitive with this 4G network system, right? Do we have to change our processes our semiconductor processes to menu phenyl sort of stuff at all? No.

Is it like to look amongst a file? What? is it? What? What's the point? is just in any any basic store. any basic silly I mean I Do Yeah, If you look at, if you look at the companies who are working on these kind of things, everybody builds them in the technology. They're the expert at writes. Alright, so if you go to a company that does, let's just say as a hypothetical this is this is not an actual example, just a variable that takes a Qualcomm right.

So go out on the CMOS There's lots and lots of CMOS chip sets and so if they were to build they would take it order. they see my make you say yes because that's what they're all their libraries are in. That's because I thought there were physical limitations, especially in the RF domain. Uh, any of substrate material.

Any any advanced seamless process that is used for making my papers is this is good enough to be good enough I Like really, you can't even really can go. You can even buy a really exotic material. now you don't need no, not anymore. Every what's the advantage of the exotic materials by silicon on sapphire? So things like not just silicon, sapphires, patil you got, you've really exotic almost nowhere.

ever seen anyone use it. So one fab in here yeah, me that? yeah and also made really small way right? Yeah and it actually looks doing beautiful as you can see right through the way. but you saw it in three five Gallium Arsenide Indium phosphate process. The advantage of those processes is the fact that they have that they kind of combine very high speed with very high breakdown voltages so you can make very high power.

PA I got like a really really low noise Alan aids and things like the law that I've been used on the front end and yeah, yeah, in RF am, right? right? But the problem with them is that they are a very expensive right? and B is that they don't offer level of integration so you can't approach that will build a microprocessor or a huge or means a little bit of yeah silicon. I'm gonna have to break it out as soon as you guys think I'm going to go on a board. Yeah, you're gonna either on the package or something area and then you lose it. And also for three five technologies, they have to be hermetically sealed.

So you say moisture is a blow. more moisture and oxygen all divided so you breath. Silicon is not. Silicon can be exposed.

Interested? Yeah, yeah, that's a Huge. You may not seem like much because it's lucky, which is a huge advantage. Even even some of some plastic packages are actually not I mean not hermetically here when you needed some kind of metallic dome over it? Yeah, yeah, so that increases the cost. Mirage an enormous amount.

So yeah, I never love integration. I mean, that's what silicon does right. Silicon is on a transistor level and in this wooden transistor level is not the highest performance thing in the world. Yeah, but you can easily beat this with some other technology.

but it's the integration that gives it the power. Well, I'm going to put so much functionality on to it, that is just you're not going to be able to get beat for example, especially for beamforming if I cannot make it power amplifier that boost out of what but I can build a thousand power a little that power amplifiers and combine their power. God and you will never beat that with anything. There's integration and in Rst eventually.

Yeah, so this is no, but military uses a lot of three Five. We have a lot of Gallium Arsenide because the military doesn't care so much about the cost and also they want in exceptional performance and things that still look into something simply doesn't give you right point five. DB Noise figure. So in figures, each cannot cannot get that.

so I not Elfin temperature ranges and things like that. But yeah, so they say they is the reason. Actually Three Five and and gas or so on advanced so much was because the U.s. monitor you put so much money into it for their communication.

Yeah, that's the other. The other thing that's really if the military doesn't care about FCC regulations right? right? they can. They can radiate into whatever like whatever power you buy or whatever with your heart here and no one cares so they can jam you. They can do it.

They don't care about that stuff, so that's why it makes a little bit easier to design. But if you want to build, if you want to build something, let's say you want to throw that 5g system is not enough to make it work at twenty he goes. Yep, just have to make sure it doesn't emit outside of your band. it doesn't emit anything in the other frequency.

This is also true for lower frequencies too, which is very difficult you have to to do. especially when the channels like Wi-Fi channels are so close to each other off. Yeah that's a big deal. Yeah higher frequencies not further the point it relax from a little because it was one will be able to build anything.

Yep, yeah so that's that's the kind of stuff that is really interesting and exciting. and then the packaging. like you said and every the integration and at the system level the optimization that takes you another step forward. Yeah so I think I mentioned this to you before at Bell Labs Did a long time ago there used to be these labs Shannon lecture series I was a long time ago and they actually restarted now.

No and so now we've had three and the first one was about the kind of artificial sensing with you merging materials with the human body and it was very very interesting. The second one which I thought I enjoyed very much was the the head of the artificial intelligence at Facebook was there and it was talking about AI and neural networks and how they handle information. because Facebook receives so much data these days I mean any almost an inconceivable amount is not not just text, but photos, photos, and videos. And again, what they need to do is that every video needs.

Every every photo needs to be processed. You know it shows up, so when you put it, it's actually even though it's real-time it's actually not writing. It gets processed at some supercomputer somewhere and then it comes back and gets posted EQ so that it doesn't have some content like I have you know pornography and so it has to be remove. So but this is not done in a traditional way.

It's done with neural networks and what is here is a learning system. Their figures are how okay. Based on other things, it continuously grows right, grows to be smaller and smaller. But it's obviously not nowhere near that.

the human brain. in fact, somebody asked this question. and the thing that makes the human brain extraordinary is how much it does with how much power it consumes. right? Yeah, it is the computation of the thoughts hundreds of the other days.

Yes, and how much it does. Whether it's just such a massive system, such a perilous system, and anyway, so he is talking about. and so Neural Networks is not like a sticky GPU for example, like an NVIDIA GPU which by the way, is used at the heart of a lot of neural networks because there is computational capability a GPU that let's say you watch a movie or playing a game. It doesn't need to be a hundred percent error free.

In fact it's not. Your GPU makes mistakes all the time. It doesn't matter care of a pixel is a rendering or either of the matter. Yeah, you really know what this is right? Yeah so in fact some of the higher grade Nvidia processors that you can buy for the ten times the cost of a regular one.

The only difference is that is being tested so it doesn't produce error marked so it's been like kinda hand-picked in a way. Sometimes not not always. but some some of them are like that and uh close those discrepancies. So I mean it is on the ancient technical question Yeah I just mistakes in the in the Mystics in the timing or some bits getting flipped somewhere or some errors that exist.

Is that because of the but it's not because of the inherent design architecture otherwise age. none of our operate identically on. Are you to area quantum? No no No. I mean another level or but it is not on it so it is a thing is about it.

If you look at the architecture of the new view itself, it is the massive computational flat level and sometimes it is pushed a little bit beyond how fast it can go. Let's just as an example, right-click you push it a little faster so it's not an overclock. Yeah, I saw that it makes mistake there. Some bits may get flipped here and there and this errors happen all the time in any computation on that unit, right? It's just that sometimes that sometimes it's unacceptable and sometimes there is error correction behind it, right? So a wireless system can can have big error rates in the order of going to the minus 3, right out of every 1000 s is one one right, which is very very bad, right? But there's so much error correction behind the system such that the overall system 0 3 or it goes below whine to minus 12 so forth.

GPUs For example, they may not do error correction because they don't care about this being slipped here and there. At the point of us making originally was that in neural networks because it's a heuristic system, actually errors the uncatted traffic Necessarily right, because in death, the system is constantly learning and constantly correcting as it goes forward. And so it doesn't have to be perfect. I Mean your brain makes mistakes all the time, right? but doesn't necessarily matter.

But your your eye is not perfect. You think you'll see in the perfect, but you're not yet. your brains in your filling in the lights. Yeah, in fact, there's just so many things that go wrong.

but it's because it's some massively parallel and so much stuff happening all the time you get kind of average is out and it turns out that as this as the head of the AI personal Facebook was talking about AI lab is that? yeah, this did. So you can take advantage of that, you can take advantage of the the algorithms that going to know. that's what. The fact? they don't have to be perfect to constantly reduce the power consumption of the system because you don't need to be so computationally lovely trying to fix everything.

It may not matter, right? So he was saying also that Ego GPU for example, you do 16 point 32 points floating-point accuracy. Yep, PC was saying that you can get it within 4 or 5. Yeah, it's that was the math right because he I mean it doesn't make that much of a difference. So I think if you're slower style very quickly.

So you were saying a next generation of DSP is so-called neural networks. don't need to be 16.4 you can get rid of all of that, bring the power down, make the system more efficient. So it's only something we've learned in recent time. Yeah, because there's that.

Yeah, it's very, very very recent stuff because experiments at that level are just simply not calm. All right, They're just basically figuring this stuff out. I Realize that you went there on not quite went down the right track, or we're pushing towards more precision, more precision. And then they find out, are we gonna train? Yeah, there's really not an actual area.

Some scientific applications you're in a rigorous yeah. An abaya for an operating system is a little bit different. For example, when you have a depth of a memory, you're putting a lot of data in and out. Let's run on neural networks through stuff.

We can get away with it so it's really good. Yeah I was a really I am I'm sure I'm butchering some aspect of it because AI is not my expertise. but you can actually go and watch that episode right? is Unbell Labs Shannon Luminary Syrian she goes I'm gonna go there on you. No need to be too easy to find.

so go and watch a listen to this. This gentleman talk about that. This is a obviously an expert. you.