Hi, this is Wayne again with a topic “How does Jim Keller get his Ideas?”.
You did it again um. You led me right into my next question: uh this one’s for me. So again I get full credit or blame if it’s a stupid question um I’ve always wondered this uh. My lay person brain you know looks at a new innovation uh. You know a new, a new generation of of Chip, whether it’s from an Intel, an Nvidia um, a tense torrent, whoever right and I look at it – and I go okay, but really tell me this this.
This idea you guys implemented. Where did it come from? You know how much of that generational Improvement is a. We didn’t think of it before versus B. We needed to try it first, but we went small before committing Big versus. We totally thought of it.
Uh, we totally knew it was a good idea, but the process – node technology, for example, didn’t allow it. We had other priorities like how much of it is a b or c I want. I want to know right, like 3D vcash, I think, is it’s all of them yeah, it’s! It’S really weird, so I mean in the platonic reality.
Everything already exists right, so we we don’t actually live in that world, there’s literally an infinite number of possibilities. Most of them are bad right, and so thanks dror strange it’s true yeah like so so that’s a problem um. So there’s one thing which is you work out? A bunch of you know: archit you’re, making a new CPU. You have a bunch of ideas.
You say these are really solid, but I want to make it’s a wider issue, but that causes you to have to go rebuild the entire cash and fetch system and then the the more instructions you fetch, the better your predictors have to be, and some predictors scale Just by making them bigger, but sometimes you need a better algorithm like the simple Branch predictors we started with were fine for years, but if you’re trying to keep 500 instru C in a reorder buffer and never flush the pipe, it has to be so accurate. It’S unbelievable now some of those things were invented. Now. Here’S a funny story which is Intel ran a competition for the best branch predictor.
They published the results and the one of them was in Wikipedia and when we first started doing Zen, we needed a really good Branch predictor. So I looked it up in Wikipedia flew over to beat the guy and paid him for a patent. I can’t make the up now the mathematical, so it turns out there’s math under these kinds of predictors, which is a little related to how neural networks in AI work right, and so the initial predictors was sort of do what you did last time and then the The better version was keep more track of the history of what you did before and use that. But at some point it started to look more like a computation of you know, there’s this space of of pass in the program.
And can you map compress that space of pass to something useful and then predict which path you should follow on, which is related to a field of mathematics and and then was it really possible to you had enough transistors, so more Zog gives you more transistors. Every couple of years – and so there’s this little and then there’s a tradeoff like in the short run, maybe I’ll make this bigger because it’s easy, but I can’t just make keep making it bigger. So I need to find a better idea or do something sophisticated and then take advantage of all the transistors like it’s it’s a combination of things and then every time you build a computer, you learn a lot about doing it. You continue to Performance model.
Software continues to evolve and there’s some things that used to be a bad idea that are not a good idea and vice versa. So it’s you know it’s it’s complicated. Let’S shift our Focus to publish stuff. You know like, like people, don’t realize how much out there there’s 100,000 people building faster computers, um we haven’t talked about Atomic semi at all and this time I’m coming in without having done any pre-briefing, because I, like, I said it kind of slipped under my radar.
Um you say: you’re focused, that’s I’m I’m at the office. That’S that’s an adom okay. I did know it was an adom, but is that the logo is that the logo for the company or okay cool yeah? So all I know is you guys are working on lowcost fabrication equipment like when you say fabrication equipment. You mean like, like asml fabrication equipment. What what are we talking about exctly but smaller yeah like a little little tiny one like that, so you want to what what your wafer is is is like 5 cm, or why tell tell me about okay, even smaller help me help me so so the semiconductor Technology is great today and all the equipment is great yeah and it got optimized to make very large wafers that move very fast. Yes right, you know they’re 12 in around, like the thing that holds it weighs 50 kg and it moves. So if you look at all the machines, it’s it’s really hard to build and heavy and all kinds of stuff yeah. You got to account for seismic stability of the land, make sure there’s no ancient burial grounds under it. That sort of thing yeah like yeah. It’S it’s amazing right, so so I met this kid. Sam Zulu, like he made, made a Fab in his garage in high school. He YouTube videos about it, and then I met him when he was in college and we started talking about what it would mean to go: build a set of equipment where you could make a fairly high-tech chip really fast, but just make one chip at a time. And then optimize the crap out of it because you’re not solving the problem of moving 50 kgs at you know 100 miles an hour you’re not trying to keep something perfectly flat over a huge surface like basically change the game and make something way way simpler and Then take advantage of like there’s hundreds of billions of dollars of material research being done. You can Atomic you can deposit Atomic layers of almost any material single atom at a time. It’S beautiful, so is the goal to be like decided to go. Sorry go for it! Well, my personal goal is to go: make a a really interesting chip really fast and uh it it.
It feels like the 3D printing said it’s yeah, it’s yeah yeah, it’s basically 3D printing. You know compared to injection molding right where you, you don’t have to deal with the enormous scale, huh yeah. Now the weird thing is: if you make them fast enough, it could be for more than prototyping right, which also happened with 3D printing.
This is uh yeah. Well, 3D, printing is amazing and it just keeps getting better and better and goes into more stuff and then there’s really fun stuff, like you, 3D print molds, and then injection mold that like what you can do today with the combination of the modern cat tools. 3D.
Printing injection molding and then CNC and all kinds of stuff – it’s it’s fantastic and yeah. So we’re we’re into building our own machines that make chips. And you know using an unbelievable amount of Material Science research. They publish everything you can buy any any any Atomic layer, deposition material. You want for almost nothing. It’S crazy, yeah super fun, okay, which um.
I I mean this: is It’s almost like a chicken egg uh question right, like which uh, which comes first for you in terms of of taking tense torrent to the next level in terms of um of taking Atomic semi to the next level? Do they drive each other like? Is this an attempt to to yeah, maybe subay real menown Fabs, you know, as the the famous quote right, I have a Fab. I computer sign company. I have ai software sure why not right yeah. Why not? What are you doing? This I mean people. What are you up to like damn boom roasted? I make I know it’s like people hey, this planet might get hit by an asteroid or blown up by a volcano man. I want to make sure that gets stopped.
I’M hustling you like they a backup Planet, but you know we got. We got big problems right here, so yeah we got to get moving. .