okay so uh we're ready to begin uh my name is Doug faar and we're here to interview Jim Keller famous CPU architect and uh uh contributor to many processor designs over the uh last several decades among other uh contributions and achievements and so Jim welcome and we're delighted to have you you all right thanks great to be here so as I said let's just start at the beginning tell me where you were born when you were born what uh your early family life was like uh what were your parents doing sort of what was the family environment as you were growing up yeah so I was born in New Jersey and my father worked at GE Aerospace and they had a program where they'd move him around so he started in Indiana and then linnn Massachusetts and I think Philadelphia and then they built a big plant in uh Valley Forge Pennsylvania and he lived we literally lived three minute drive from the Valley Forge GE Aerospace office and I was one of six kids the second and um my parents my father was a mechanical engineer he designed satellites my mother was salid Victorian of her college she had been a trained to be a teacher she's very smart um but she raised six children and she was not working when I was a kid for the most part later she became a therapist and uh you know was successful at that but as a kid she was a stay home mom and I grew up in a house full of kids there was six of us relatively close together so was the uh your fact that your father had an engineering background was that an influence did he work with you did you do projects together did you anything like that yeah my my father was very good at describing and drawing anything so I grew up in house where he'd come home and say oh I saw this bridge let me sketch it out for you and he' quickly sketch out the whole bridge and then tell me why it worked it didn't work and I I sort of grew up thinking that's how things were and my parents were both members of a book club there was books all over the house and you know they had lots of they believed in intellectual conversations uh they were pretty smart people generally speaking and um they had a family friend who talked to me on a a family trip like the the kids swapped cars and apparently I talked to them for hours and and my father's friend told him that I was one of the smartest kids he'd ever talked to and and the reason I mentioned this because then I was dyslexic apparently you know later diagnosis and I didn't really learn to read till third or fourth grade and my parents never worried about me so they thought oh he's fine he's just figuring it out his own way and I think in the modern world that would have led to panic attacks you know your son in second grade still can't read and and but it never impacted me I I I noticed in second grade that I was in the group with the the books with the really big letters instead of the smaller letters and you know it just didn't impact me very much and then periodically I would learn like I remember in the end of third grade I learned to read and suddenly I could read books and books are fun and I read a whole bunch of books so it's slowed you slowed you down in terms of learning to read but you became an Avid Reader I became a reader never a writer um I struggle with writing like I can talk mostly a Blue Streak but if I if I can see the whole page of words I can write them down and I think maybe in a world of text editing and writing and rewriting it would be better um but as a kid like writing was always very painful um but but math was also they call it episodic learning like I remember we're learning long Vision I didn't get it like and then one day it was like oh they're guessing right and it was like bang I got it and then the same thing happened in algebra like f ofx equals something I didn't know what they were talking about like I think I went the whole year of algebra just thinking I don't know what they're talking about it was like oh it's symbols and and then I just got it and so I I I'd have these moments but I I was you know doing my own thing as a kid and you know like school was you know time off from doing other things and uh so you have five other siblings that any of them follow a technical track my older sister was a mechanical engineer uh my younger brother is a computer scientist and very very good and yeah my my siblings were all pretty smart but there's a a therapist a sociologist a philosophy Professor we did fairly different things and interesting family so you're progressing through High School anybody in high school particular teachers or whatever that kind of steered you in a particular direction or got you interested in things yeah I I had a couple really good teachers so I went to a fairly large Suburban High School in the Philadelphia area um we had a math teacher who who taught calculus and for special presentations he'd put on his good suit and his brightly colored tie and make a really big deal out of it so the fun fundamental theorem of integral calculus was a an event you know and there was multicolored chalk and you know he was he was a he was a character um and some teachers in our school were were basically terrible and we didn't do anything and in our school we we sort of run a m we like attendance was optional and uh my friend was a student class vice president our president and our class made lots of money on concessions for football games and we we kind of saw ourselves as kind of operating outside of the the school um it was kind of a wild place for a kid um the the principal subsequently was fired and by my brother said I went to a way different High School than you did um to reing these students in huh yeah so that was you know the end of the 60s was pretty tumultuous in this country and you know I grew up you know watching television to find out if my uncle's had died in Vietnam or and it was pretty traumatic I'd say and then the 70s was kind of the the after masses when I went to high school and and it was both good and bad I had like I said some spectacular my geometry teacher was great my uh one two my my chemistry teacher I still Mrs Anna fko I still remember teaching Us St geometry she was fantastic teacher and and then other ones were like they didn't care we didn't show up no seemed to notice and I graduated high school with bees with literally no plan except for a girl I liked up the street was going to Penn State so I applied ah that was the uh the path to Penn State huh well yeah there was a funny event I got called by the student counselor like in my senior year back then like like somewhere at the end of your senior year somebody might notice whether you're going to college and nowadays they start prepping inary school years ahead of time yeah so and she said you're a really smart kid you should go to MIT if you ever heard of MIT I think I vaguely had heard of MIT um I thought nah I don't want to move to Boston and then she gave me a list of jobs sorted by salary and at the time the top college graduate salary was chemical engineering then petroleum engineering then mechanical and then electrical and I thought the first two seemed dumb so I flipped a coin between mechanical electrical engineering but my grandfather is electrician and as a kid you know when we visit him his house was full of electric motors and washing machines and all kinds of crap and I thought it was amazing and uh so I randomly picked mechan electrical engineering because it was high paid I drew a line cuz below the halfway line was biology physics oceanography like you know if you'd ask me what what I wanted to do in college every single one of them was below the line but I thought i''d be done to go to college and not have a job that could pay any money so and my grandfather's electrician and I thought that would be funny so I I signed up for Penn State I got accepted electrical engineering I called my grandfather and he said watch your power factor which is the angle between current and voltage and electric motor so that was my that was pretty much the whole story about you know going into engineering how you get there so when you went to uh took was this uh was was it electronic micro Electronics Focus was it um what was the focus of the education I had one for training course I had one logic design course I um Penn State at the time this is 1976 uh they had a very good dou Department um it was like six six or 700 students is Big um yeah big and they started the Core Curriculum you know your first year was math physics electives I was uh also a dual major in philosophy which was a funny story and math was great mechanical engineering was those were like the weed out courses and I remember learning to solve accelerating rotating reference frames and I thought that was amazing and I really I really liked College I uh i' say the first year was um it took me a while to get up to speed because uh you know high school we just messed around all the time like why I had good teachers we didn't have to do any serious studying and the first year was and I started out strong but by the end of the year my grades were in the pits and my my father said if I wanted to fail College I should really think hard about that in the summer because rather than waste money on school I could go get a job and I thought I'm not going to fail college and uh so I did pretty well the rest of the so how did you get into philosophy uh I read a lot of books as a kid and um I was really interested in like ideas in general and so I I was a double major and then the the funny story about that is I got a like in my uh second year I got a letter a note s to the electrical engineering department from the head of the philosophy Department that they wanted to meet me and I went in there I was kind of excited was like wow the head of the department called and um he said we've never had a student like you like literally like there's no double e with a philosophy major and and then I just wanted to show you something he shows me a paper written by a philosophy student which is like he said this is a midterm exam it was like five pages nicely written double spaced both sides and then he he pulled out my test result which was literally a half a page with multiple scratchings out he said I I can't even read it Jim I don't I don't know what you said and um he said we write a lot as philosophy you know students and I check with your teachers they think you're great um but I have to tell you you'll never get a degree in Philosophy from Penn State University but you can take any course you want you'll always get bees so I got kicked out of philosophy because I couldn't write a paragraph in a in a test environment so it was pretty funny but I took lots of classes like and uh it was really really an interesting thing to do so uh well that's pretty interesting story what uh so you finished up there uh you have any idea what you wanted to do afterwards what how did you yeah so Penn State I started I really um I didn't really start electrical engineering till the second year because that's how the curriculum worked I got straight A's in dou e um I I got a call actually after the philosophy Professor from the head of the department in double e and I thought oh I'm getting kicked out of dou e as well like this happened twice when he said I just wanted to meet you we don't have that many people get straight A of Penn State and I said well I'm working really hard at it and he goes that's obvious so the the thing I really like was electromagnetic fields but then my adviser was Dr Joe stack who's head of the semiconductor group and IBM had donated to Penn State a 2-in wafer Fab so my senior classes I took a bnch yeah little tiny things well we laid out transistors on Vellum with tape oh yeah I know you know like six Micron transistors they were big and we made transistors in the lab and I still remember the safety briefing on HF and so so my senior courses were semiconductor focused my uh you know the the thing I thought was you know I was going to be electrical like RF engineer um but that was really interesting and I took a programming course which I thought was really fun it was punch card Fortran and I had a logic design course which I thought was curious curious um it didn't go very deep but it was amazing that you could build anything out of something so simple so electromagnetic field the is is complicated and almost everything becomes almost incomputable right even with relatively simple antennas with relatively simple driving things it becomes very difficult and nowadays we built really elaborate simulators for that kind of stuff so the simple equations are cool you know but the the application quickly runs into the computational wall and Material Science was worse it was my favorite class in pet State because the professor was great and the descriptive language of Material Science was amazing but in the books all the problems became curvefits to simple equations because everything outside hydrogen atom is insolvable and it was really an interesting again as a philosophical basis for science it's it's a very curious thing how big the the leap is from you know the foundations of physics or something and the application of it and then here's logic design which is the leap from logic design to the application of it is a carom map which is Trivial and then you know it's it's really kind of wild so I wasn't really sure so I interviewed for a whole bunch of jobs mostly and semiconductor companies Fairchild Motorola but I randomly did an interview at Harris in Florida and I thought I could literally live on the beach and go surfing and they had a really cool computer project that was doing fiber optics combined with writing data on film at high data rates and and they needed a logic designer basically and so I I took that job and so between the environment and the job seemed like a good place to go yeah and then the boss was a reader and he had a bookshelf full of books so I'd written it read as a kid and we had a great discussion and then it turns out he was subsequently fired after they hired me and when I showed up they literally had nothing to do with me they were like why are you here I like I I I had a letter says show up you know August 15th and like well the college students all started in June it was like yeah but I drove across country back and I negotiated that with the recruiter and they're like we don't know they gave me badge and a desk and then after a couple days somebody wandered around and said we have a whole bunch of test equipment to fix so then I spent about a year and a half fixing stuff and it turned out to be great I so you never got close to computer design uh there well they had some boards with microprocessors that didn't the work they had a power controller for Florida Power company that would remotely control that that didn't work they had a whole bunch of random projects and I was a busy so were these projects that somebody had designed but still didn't work they had couldn't figure out how to yeah I got put in a group or they had multiple like basically logic design problems to do and I did a little bit of logic design but mostly I fixed stuff and I did one high frequency board at the time 40 mahz STL board that encoded data for a fiber optic transmitter that didn't work so that was like my first electrical engineering job where we had to fix the ground plane and a whole bunch of stuff to make it work and then that job is why I got hired at digital equipment so how did you find your way to digital equipment you're down there surfing in Florida and there was an ad in E times I thought I I wanted to move back North my familyes in Philadelphia Boston seemed Close Enough by so I read the baky E times and there was an advertisement for uh a clock designer at a computer company I'd never heard of digital equipment so I I got an interview at Ron the day before to practice because I hadn't interviewed for a while and then I and they gave me an offer on the spot which I turned down and they were pissed about it because they paid for my plane tiid for your trip yeah it was pretty funny and then uh they're like did you have any attention to working here I said yeah if you guys weren't you know weren't weren't being like you were so it was pretty funny and then I went to digital um where they needed a high frequency clock designer and they quickly determined I I knew about that much about about it but they hired me anyway and U and that was great I was that was so what group did they stick you in so I was the vaccinated 100 group so the lead architect was Bob Stewart and he was the architect at 1140 1144 117 780 and 8800 so he's one of their best computer Architects and I work for a guy doing the cach sub system for that computer doing logic design and um yeah I told so you didn't know anything about really about computer architect I cashes or anything like that so A friend of mine gave me the 1170 and 780 manuals like the day before the plane flight and I read them on the plane and I basically went in and said I every single person I talked to I said I had a lot of questions and I I basically had the manuals and it said why' you do this and what's this for and you know I was um i''s say fairly confident as a kid and um and they basically said we don't know what to do with this kid but he's we they'd never seen anything like it apparently so so I got a job there and it was great I worked for that group for about seven years and on the uh 8800 the 8800 and then there's two follow on which unfortunately got cancelled one for political reasons and one because the technology didn't work out but I uh I did quite well doing logic design and then board design they were ecl computers so there was electrical design board design logic design and somewhere in there I started thinking about computer architecture and CAD tools and I wrote a logic simulator a timing verifier as fairly large software projects you know kind of at night literally because it wasn't my job and then did they have anything like that or yeah we we had bought there was a startup in Silicon Valley called ballad they did a drawing system that had associated with a a timing verifier um it worked okay but it was difficult to use and then digital had a simulator group that was very thorough but very slow and at some point along the way so I worked for Jay Grady was the manager but Bob Stewart was the architect and Doug Clark was another architect and these guys knew all the digital had a fairly big research community and they introduced me to people like Forest basket and Butler Lamson and you know and then we're explaining to Butler the logic simulator work and he said that's crazy here's how you should do it and he just sat there and he he stood up and wrote down on the board here's what you should do if you do this it'll be way faster so I wrote a logic simulator that did exactly what he said and called him up and told him how much faster it was he said that sounds about right good work and uh that's per that yeah and he was uh yeah I don't think he did a lot of implementation but he could think really fast he told everybody how how it should be done and everybody went off and did it like you yeah yeah and it was literally exactly what happened and um and I also got to meet Chuck attacker who was another great architect so I and they kind of kept an eye on me because I was a little bit special and um but I got to meet those guys when they were yeah they were at zero Park when I was there working on the Al and all so yeah so Chuck and Butler were system research lab and then Jeremy Dion and Alan Eustace and a bunch of guys I met were at the Western research lab which are literally across the street from each other but they were like the two factions of research at digital and I somehow got to know both of them did you work with I Dan Dober poool yes was he involved directly or how was what was your relationship so I was at digital about 15 years the first half was in mid-range systems Group which was e design big computers and then I joined the semiconductor group in about 89 or 90 so I spent six months at Western research lab where I met Norm ji and then worked with alen Eustace and a couple people on the bips computer which was out here out in silic Valley yeah in palto and then when I went back Pete Bannon took a job in the semiconductor group and he said you should come here and we became co- architects of ev5 which was the second Alpha chip so Dan did the first offit CH ev4 and we did the follow on and then Dan at some point moved to California and started the the arm team for for digital which was a strong arm team so I I knew him pretty well we talked a lot but I didn't work with him directly at digital he was one that we tried diligently to do an oral history with and never succeeded before he passed away yeah um so I was curious about your remembrance of him and what his uh yeah he was he was kind of amazing so yeah they when the first Alpha chip came out it was you know 166 and then 200 MHz which was literally three to four times faster than any other clock frequency of a processor ever built before and Dan's desk was always about this big and it was covered with schematics about this high and it's like how do you find anything he said if you don't touch it I can find anything and he would pull out a schematic from like one year ago and say Here's how you do this here you do this and yeah so getting back to uh your projects initially on the ecl 8800 right the uh what were you you know what was the the biggest project that you were responsible for did you lead a certain area or what was your yeah so so in the v800 I was literally a board designer logic designer so I did the bus interface and I built logic to Specs that other people did and then I I started working with one of the teams that did the performance model architecture because I like Pete and I became the guys who debugged the computer Pete Bannon so like he was a micro coder I was log designer when we powered the machines on the lab lots of people are there but we basically lived in the lab for a year and debugged everything and ran all the software and as the software started coming up the people did performance modeling would want to try things and run it and then some things work like we thought and some things definitely did not Pete rewrote a lot of firmware and then in that process I learned a about how computers were actually built end to end and the guy who was doing like one of the io boards quit and they needed somebody so I just went through the whole schematic set and became the guy who debugged that and then we delivered machines inside the company like the first 10 machines that were installed in the operating system group you know as soon as they brought up the software and they started loading them they started to crash and we debugged that and so what was your what was your learning what did you take away from those few years on the 8800 project in terms of later later directions I mean your the you you became pretty familiar with the whole architecture of the machine what worked what didn't you know did you did you come away with any opinions about this is the right way to do it this is the wrong way to do it so you need better tools you know what were what were the takeaways from that project yeah so at first it was just a whole bunch of parts you know like here's a here's a piece here's a piece here's a piece but then it started like Bob Stewart built you know he built basically a risk pipeline in a cisk machine early so we had a decode stage you know like instruction fetch decode you know register read execute data cache look up right back so the Vex 8800 had a pretty clean wrist Pipeline with extra stage or two in the beginning to decode the the the CIS instructions but executed them as Micro Ops so you know first like he architected I didn't know what that but by the time we finished the project I knew what pipelining was you know I knew what the cost of Branch Mis predicts were cuz back then we didn't have a branch predictor so when it went the wrong way you just flush the pipe um what a cache was so basically it was like architecture 101 for three years and by the time I was done I how caches worked how branches work instruction decoders faxes were kind of complicated because we had multiple levels of priv privilege architecture um they took interrupts we had micro code some of the heavy instructions so how that whole worked and then the following one was Argonaut which was take the basic architecture and double the clock rate um from that I learned we pushed the clock rate further than the technology wanted to go it made it really hard to build and also that the tools to do the analysis were really terrible um so that's when I got involved in writing cat tools so I WR like literally at first it was timey verifier and then the the simulator which we then used on multiple projects and by the end I hadn't written the performance model but we did this performance project where we took one of the boards of the VAC 8800 modified it to record all the micr PC addresses and some other information and then Doug Clark wrote a paper which was here's how a vax computer spends its time and I contributed I built the hardware for that Pete wrote the micro code Doug did the high and little analysis and I was sort of like translated all these moving Parts into a computer with pipeline features to a computer that has performance because of how it spends its time and so when I joined the alpha team Pete and I wrote the performance model for ev5 from scratch bar cells and you know that was our first experience on doing that as a computer designed from an architecture spec and a performance model first that translated into a chip you have during this time is there much interaction with the com compiler group and some so the compiler group was very focused on high quality compilation so there's an interesting Journey so during the CIS era at conferences people would show up and say my computer's better because I have more instructions right and then the compiler era was our compiler is better CU it's reliable it's solid it doesn't have memory Le like there was this whole thing and then the risk gu showed up and um they built Titan which was a risk machine at Circ at system research lab sort of at the end of the vax 8800 and it was half the boards a quarter of the volume and faster than the vax 8800 per processor it was also simpler it didn't have the privilege architecture it wasn't a multiprocessor and there it was like this computer is so simple it's easier to build compilers for and the compilers are contributing to Performance and that was a really big transitional moment because the compiler up until then weren't the guys delivering performance they were delivering functionality quality and when we transitioned to Alpha this the alpha team had an internal research group that did compiler performance work and then the compiler group at some point created a pipeline from performance ideas into production compilers but that that took years to do it at digital and ultimately very successful so your observations of the Titan and the results that came from that later influenced sort of your own Direction and sort of being able to sort of see both uh yeah it's it's kind of hard to say so so Bob Stewart believed in making Hardware as simple as possible and Titan was was definitely up that alley but when we got to EV like ev4 was dual issue we'd call it super pipeline today ev4 ev4 was the first Alpha chip okay and and then ev5 was four wide a little bit more pipelining had a multi-level on chip cache which was probably a mistake um and but to fill up a four wide in order machine you really needed the compiler to work and then this is where the debate started about was compiler going to generate more and more performance or was the hardware going to do it so ev6 was the third Alpha chip Dirk Meer and I architected that and that was our first out of order machine and the premise was software could take you you know let's say Baseline performance a 1x with compiler improvements you get the 1.5 and some people said well we'll just keep improving that but the search space inside the compiler blew up the VW projects by and large did not succeed in terms of delivering performance without unbelievable investment and out of order became the way to do it but the when we built ev6 there was still debate about whether you could build a high performance out of order computer make it work and some simpler ones had been built but ev6 was four wide 20 instruction window 100 instructions in Flight like it was way past our capability to think simply about um nowadays that would be considered fairly simple um but that was a big transition and by then I would say I was you know this was back in like 96 so I've been working for 16 years yeah so you were deeply embedded in and I would say by by ev5 was my first big swing at an architecture that Pete and I did we learned a lot when it was finished it was both the fastest computer ever built or microprocessor were built and I was let's say so embarrassed about the mistakes I I made in it I could barely talk about it like it was really uh a funny personal experience because I knew every single thing wrong with it despite the fact that it was actually pretty good and the team that built it was a very good team and it was interesting so for ev6 we rewrote the performance model infrastructure from scratch Alan eustus I I described the performance model he wrote it was like 30 th000 lines of code all kinds of complicated stuff he said Jim I wrote a performance Model A th000 lines of code like what what why is there just so much code like literally what did you do wrong and it was like it was like a a gut punch like how did I screw this up so badly so I wrote the uv6 model in the first version of it was like 1,00 lines of code and I basically stayed up for a week getting it to a th000 lines where I could actually parse the alpha architecture all the traces run the out of order model inside the or not the memory system and made it work in a th lines of code so I called it out and said I did it thousand lines of code I said I said I still don't know how you did he goes oh Jim I never did that but but I knew how competitive you were I thought if I gave you a better goal he said 30,000 Lin seem crazy to me I was like that's pretty funny but it made me really rethink exactly how you build performance models like really think deeply about it which before it was sort of you wrote a model explore ideas as opposed you you wrote a model as a thing that has value that is understandable and clear and usable and then Durk and I went back and forth like like he take the model he say I'm going to go add this and this and this and he'd just go rewrite that stuff and he was a much better programmer than me and so the ev6 was the end of the road there at at deck yeah so we for both you and the company I presume so when we taped out the chip Durk went to AMD where he ultimately LED K7 and then became CEO I stayed until we had Linux not Linux Unix at the time running running on multiple chips and then I I joined him to work on K8 and then they did a a follow on called ev7 which was the ev6 chord with some tweaks but with basically onchip memory controllers using rambos and they built a 128 processor machine out of bv7 and I think that might have been finished after compact bought the company yeah so yeah about three months after I joined AMD compact bought digital equipment but Pete Pete built ev7 and stayed and finished that computer so then so you went to AMD what was the uh and your friend was already there and recruited you to to come and work I work for Te RZA who was O So AMD a couple years before had bought NextGen and their processor the 586 became k6 so AMD basically 286 386 486 were Mass copies of Intel processors K5 was the first designed by AMD processor and they made it work but it struggled they bought NextGen cuz they had the 586 in production and then they tweaked that to become k6 Durk led the K7 project which I spent a lot of time on and then I I started the K8 project so what was the goal of K7 and K8 just the Next Generation x86 uh so I'd say K7 was architecturally looks somewhat like ev6 that Dirk and I had built but it had the macro instruction architecture for cracking x86 Ops and at the time we thought you had to so in a risk machine you have a load instruction a store instruction add a branch instruction they're all separate instructions but in x86 you could have an instruction with a load operand register operand maybe a store oper hand and then some instructions dive down in the micro code and there was a there was an argument at the time of should you fetch a macro instruction and take it apart or should you take it apart and issue Micro Ops and K7 on the iner side was a macro instruction architecture but the cach system and the bus interface was literally licensed from digital was the ev6 bus and then on K8 the design we started that I was involved with we were going to make it much more of a micro op machine all the way through and then we also decided to do 64 bit so my big contribution to that architecture was I was one of the principal off authors of the X x86 64 spec the hypertransport spec which is the glueless multiprocessor spec and a a big update to the chip architecture that ultimately became opteron now the I left to join startups during startup you know craves in 1999 yeah so were you working for atique during this time at yeah so I was working for atiq and then atiq left and then I left to start a company with a with a friend being funded by atique and then he turned out he couldn't do that project so then I joined and arra pole at cite so what was that company that ATI started uh we I don't even know if we got a name so te was in the mode of funding startups and we were going to go build a network processor and build a CPU so so AMD I was there for three years and we we did do the 64-bit spec we did the hyper transport spec uh we art laid the seeds for the what they call you know multicore chips way back then and that mostly went pretty good so cite you went to cite and Dan was Dan dool was there right CEO so he recruited you there yeah and and then I was that was very that was a different kind of processor right yeah so we did there was a myips processor and then we built a networking s so so it had gigabit Ethernet controllers memory controller processors and some accelerator functions and we talked to literally hundreds of network companies at the time and um that's where I met Andy Beal shim and a couple other people at Cisco and they basically laid out like literally the napkin drawing you build this process or you know Cisco will buy it and we built that chip it was super fun it was like a really small team and that company was acquired by broadcom and we were there other deck alumni and so forth that staffed that where yeah it was a mix of a deck team there was some sun people and then a whole bunch of random you know contributors we had people from Intel all over the place and how big a team was that I think it Peak 125 130 and then I was there for we were there about a year and a half at cite and then between that in broadcom I was probably be there for four or five years so getting back to AMD what was you've now worked on a very different sort of instruction set but you said you copied some of the architectural learning from deck what what were your takeaways after AMD what was you know you talked about all the mistakes you'd made in the previous Des there so many so so I you always learn from our mistakes right yeah yeah so I gave a talk at microprocessor forum for about ev6 which was super fun because talking before me was the power four guy uh one of the PA guys from HP and um and one of the sun guys and uh and the IBM guy was especially funny he goes man I have to talk after you guys I can't believe it our processor is terrible so so power four was really cool architecture but the frequency was low and it was built on multiple chips on power was whose uh IBM okay and um and then the HP chip was really high power and they had it was really interesting architecture sure and and so I went out and I I started my talk I said I came here to talk about the world's hottest you know biggest and fastest chip and I just found it's only the fastest cuz the IBM was unbelievably big and the HP one was unbelievably high power and um but it a digital we literally designed down the transistor level like the design itself was 150 schematics and almost every one of them we had some called macro block which could be a collection of transistors that you would Stamp Out multiple times we had a small cell library but we figur out one time there was 26 different flipflops in the chip but a lot of them were just handbuilt by people so and we used to joke that we had both things in our library and devices andp devices and so optimized all the way down trans and it turns out that's completely not scalable right and at AMD they had a really good sandard cell methodology where they could write RTL and then translate that to Gates but lay the gates out in the way that you didn't lose that much performance but the gates were well characterized and that was sort of amazing to think about let's say how many obstructions levels are there in a computer because at the bottom you have atoms right and then you have transistors which are dope devices and then you have metal stacks and it at digital we're designing at that level but then you have standard cells and then slowly you have C had tools at Place cells and then you have RTL with synthesis and the let's say understanding that computer design was a a layered set of abstractions that is fairly complicated and each one of them matters and at some level each one of them trades off some something but lets you do something bigger and yeah that was a big event I'd say so you found the methodology did did AMD have the tools uh some tools were pretty good some tools were less good uh they rode a lot of their own placers uh a friend of mine said he felt like he was like a human sythesizer so you take the RT and then literally have to draw the gates for all the critical pieces nowadays you always let the synthesizers do the work um yeah it was it was quite the quite the journey on that stuff so did you're getting now moving ahead to cite did you uh was all this directly applicable did you have to learn new stuff to kind of going to a network oriented processor at cite we still so the deck gu still really believed in custom design it was probably to Custom Design and but very friendly with the transistors and like we literally first name bases with the yeah we made our own PCR Express FES like the idea of buying libraries and FES for people was just kind like was considered crazy um um but we delivered the product and the thing I think I learned at cite is I talked to so many customers and saw so many network devices and and a lot of them had secrets you know like like they're doing this new special network box and I would literally talk to them for two hours and reverse engineer from what they said everything they were doing and that gave me this kind of let's say meta understanding of how like that kind of system design goes together and then at some point I realized how smart like the chip we were building was actually partly direct instructions from some senior people and partly talking to lots of different people about how to make a part they all wanted and the embedded business is is curious because if you put everything in that everybody wants it'll be too expensive and nobody can afford it but if you miss important things nobody can buy it because they can't use it and the the trick in the embedded world is hitting that sweet spot between you know enough stuff that makes it a great part but not so much stuff that makes it like stupidly expensive and that was a like a really curious thing to do is you know it's almost like a sideshow from the you know you think you know going from digital to AMD the like in processor design that was one path and this was like like a different kind of system engineering but it turned out to be really really interesting to me so it was it was quite a lot of fun so you learned a lot about applications stuff in and then at Brom I met Henry Samuel he was a another genius guy and I I traveled around with him one time and we go to talk to new customers and like I would do my pitch on side bite and I also knew about two other products and he would listen and about the third day he said let me take this one and he gave a better talk about our part after three hearings and I had done like he was kind of a wild character um so you worked at uh you know big companies deck AMD now sbes a startup basically yeah and then PMI was a startup yeah we'll get to that in a minute uh was that did it matter to you sort of your well my so the VX 8800 project and the alpha projects were both small teams so the companies were big but the teams were like 100 people and so there's a certain Dynamic about how teams worked for a while I had a theory and I had some data which was like you know takes 400 people four years to build a computer and 300 people 3 years and 200 people two years right and 100 people a year and a half and I wasn't sure a team bigger than 400 people could ever finish one right but then itanium was 800 people and they finished a computer so I thought oh that's weird CU but it turns out a certain group of people inside the team like literally detached themselves from the management layer and built a computer like I talked a whole bunch of of people worked on it and they sort of finished the computer despite the fact the team was too big to actually build up something and AMD you had a big team right or is that later the second round of AMT you had a well yeah when I went back to AMD the CPU team was 500 people but they were doing multiple projects they had uh uh you know their bulldozer core the Jaguar core and then a couple of teams to do shrinks and then put the processors into the console parts so and then there was a fabric team yeah but that's the second generation AMD right so let's get back to broadcom what was the uh you you got a whole brand new education in terms of different uh application space for these processors right yeah so Henry Nicholas believed in really owning a platform so you going make a network computer there's a data plane part uh control plane part there's nexx there's all these different pieces that go together and one by one they don't add up you know they're not real products but the whole system is a product and then the trick was how do you build that that actually makes multiple people happy and then yeah it was really interesting how much detail there was and then networking at the time was split between the ethernet world and the Sonet world one was cheap transport one was guaranteed bandwidth allocation and guaranteed latency and then those features over time kind of fused together and networks were also built on abstraction layers the famous seven layer Network stack and yeah it was like it was a great engineering you know tour to force to be with some of these guys and work on it all and and then there was this kind of curious thing about the physics so when I started you know 20 mahz like 100 mahz was a really high frequency I still remember seeing the fir seeing the first 10 gahz scope I was like what the hell happened like how did we go from 100 MH was hard the 10 gigahertz as possible even and now we have 112 gig Sardis and and it's again there's so many layers of what makes that happen and I can kind of see them like I know how packages work in wires and wire bonds and inductors I I used to be a you know electromagnetic field person and it's just a miracle like and and they literally watch it evolve over time and I was talking to Henry Sam you know like when are we going to have gigabit Wireless and he basically did the math like he literally just sat there and did the math well here's the slope of the curves for transistor performance and here's the size of devices and geometries of packages and here's the investment and you know if you want to do it here it'll cost a trillion dollars but if you wait five years it'll cost a 100 billion and here's how the curves going to go and I predict 5G and you know 2018 and that's actually when it happened like if to him it was like well these These are shapes you know and the push bandwidth you know there's the analog signal bandwidth but then there's the encoded like symbol bandwidth and that was a space like that was a design space and then there's parameters underneath it moving like here's here's how semiconductor is moving here's how packages are moving here's how encoding works and he had a plan it was like was like you're kidding and um this is the future and this is how we'll do it and this is well you know the famous you know Mo's law so when we built the V 800 and the following we didn't know about Moors law MH like inside digital like maybe Gordon Bell knew or somebody but like we made the VX 800 F faster than the the vax 11780 because the customers wanted more performance but there wasn't like here's the curve predicted byor law and if you're above the curve you're going to make money and if you're below the curve you're going to lose money like that wasn't the thing in the 80s and and then the is this people but not for the system people yes well and you could design a computer like the backx 800 was designed by computer Architects using Gator a chips and a couple of custom chips as components to achieve our goal to build this computer whereas in the Alpha Team it was like they built transistors and like literally needed to do something with with them and the mission of transistors by then was to make the transistor smaller faster put more on a chip and then you know Dan's you know genius observation was if you build the pipeline the circuit technology and the clock networks properly you can make the transistors run essentially full speed and and then you don't push it too far on pipelining because then the logic design breaks but if you get the clocks and the flops and latches and all the tools right you could make 200 MHz processors in uh like literally half Micron Technology which like you look back and you go what what happened there that's that's kind of amazing so on to PMI that was a pretty big transition going from well it started literally as the reboot of of cite okay so we got uh architecture license to power PC which is long process and then who who were the other found I mean were you a Founder was it you I there was the original three founders from saite who were the senior Founders and then five more of us that include Dan yeah Dan daro am Gil Leo Joseph were Founders and then panit Kumar uh Mark haer myself Shri ball and Sanam were kind of Founders but I think we weren't really senior Founders or like the docks we're definitely I understand that right let's say hierarchal and so the original idea was let's go do a similar part but upgrade everything so 1 gig went to 10 gig processor performance went up it was an outof order power PC core and it was a much more aggressive design in the architecture so that we could stream to make the performance work you had to stream data in from the 10 gig ports go into an non-chip cure memory process it with out it having to go to memory and then be able to do interesting things and send it back out so it was architecturally much more aggressive uh part and then at some point it was also architected in the way that you had a processor area you had a slice that had memory controllers PCI Express and network stuff that was pretty generic and then you had the network acceleration stuff and we built the part so you could tape it out two ways one as a network processor for you know the Cisco business and the all those guys but you could also build it as a standalone part which could have been a you know a personal computer part for apple and we talked to Apple quite a bit about so I thought that a major goal of PMI was lower power was that did I make that up or um Dan was in the low power um because he he did the strong arm project and then all the network guys they had lots of boxes with limited air flow for for 25 watts you could put a part on the board and just blow air over it for 50 Watts you had to put a heat SN put it in the right spot for for power above that like things got exciting pretty fast and then there was some low power circuit designs and tricks and the team was pretty good at clock ating and low power high low power but very accurate clock distribution which I think was better than industry standards um by today's standards it wasn't really a low power design but but for the time it was actually pretty good so the goal was performance but with minim minimizing the power along the way yeah in the network processor world it was sort of performance but hitting certain PowerPoints so you could deploy it properly right like as a system engineering problem right um the PC World is much more fine-tuning power performance by then so desktop parts are 95 Watts because it's easy to put a fan on them you know like desktop is go fast as possible as soon as you go on the mobile now you're trading off hotpots on the package and battery life and at the time the mobile parts were starting to get plausible they were still 25 35 Watts so was there did PA semi actually end up selling parts or you what the company was acquired by Apple before it actually offered we taped out a part that worked but it was purchased and you left semi before or you went to Apple first and then yep I I we finished the part taped it out and then I joined apple and then Apple so then there was a funny situation where Samsung was partnering with apple to build a 1 gahz arm processor MH and a guy I'd work with so Samsung partnering with digital equipment and they built What's called the EV 56 and 68 Parts which were shrinks of the apple chips on the Samsung Foundry and the guy who ran that project called me up for my helped to get the One gtz processor to work and I said I already work at Apple uh but you should call Dan and Dan will help you get this because they needed they needed a big client let's see and then Dan figured out what was going on he called up Steve Jobs and said why should I help Samsung when you could just buy my company and we'll we'll build it directly and and apple acquired psme so how did you wind up at Apple I had some friends there in a research group that said you should come do something fun I said what do I work on they said you can't we can't tell you and then for a little while I think I went through at one point I was I walked through three locked doors to get to my office because Apple loved secrecy partly for the drama I think and partly for uh Steve didn't want this project messing with this one and everybody like you do your stuff you do your stuff you do your stuff so were the people that got that got you to Apple were these from digital or yeah Dave Conroy I work with he was at the system research lab working with Chuck ther and he had joined Apple to be one of their like Guru computer guys and then I met a couple other people there I'm trying to think of who I knew besides Dave I don't think I knew oh Steve Paul's was another guy from the digital system group he worked in Mac engineering so I knew like literally three people when I joined Apple so what was the state of processor design at Apple at the time I forget exactly what that this was yeah so I joined so the first iPhone chip 2009 yeah something like that it 2009 2008 yeah so the first iPhone was shipping there were literally weeks from taping out the second chip with which was called H2 internally and then and that chip was the F the chip of the first one was a arm-based yeah the Arm based arm IP the first chip was literally the person who said he was the architect like he showed me the architecture document which was an Excel spreadsheet with all the IPS he needed so he said I need this CPU this memory controller this bus this you know it was just together and so they gave that to Samsung and Samsung put them together on a chip and it worked and the memory latency was terrible and performance was weak and they could barely get the software to work on it but uh software guys figured out how to make it work they did a an unbelievable job because they took basically uh OSX and shrunk it down and got it to run there and by doing that they optimized a minimum ize the software so later on when we gave them faster chips the software could really grow like somebody said the biggest difference between IOS and Android was Android got ported to a chip when the phone chips were already pretty good and it never really got cut down in the way iOS did and and that may be true so I I got involved on H2 and then Tim laay and I were like co- architects of H3 which we really optimized a lot and H4 is the first chip where we put a really big Graphics unit and then H5 was the first time the PM team delivered the first processor called Swift and then H6 is when Gerard Williams and I designed essentially the big out of order computer that Apple's been building on ever since so before that one were there was it a matter of executing on sort of known architectures and so forth and just getting the power out and so forth to make it work in this environment was well the original phone so's were pretty straightforward it sort of like you have a radio interface you have a display interface you have a keyboard interface you have a processor that runs the operating system there's a little dpu to to render the screen like like the diagrams were you know a piece of paper this big with a bunch of blocks and interconnections so the the architectures of phone chips have mostly been straightforward since the beginning the thing that makes them let's say magical is really hitting performance per watt and power and then Apple was very aggressive on new things so we put 64 bits in before anybody we did high resolution display before anybody uh we put a lot of graphics performance in H4 like it was just completely outside anything anybody ever done like if you looked at the chip the GPU is literally five times bigger than the GPU in every other phone chip and some of that came from Steve Jobs and Mike hbert and you know Steve had this idea to make the best technology you need the best components you know or make the best product you need the best technology best components and then the Apple guys when they started really getting into touchscreens and visual apps they went hard on stuff like high resolution lots of graphics performance graphics performance in service of the computer interface as opposed to graphics performance in service of just a game like Apple's never a gaming company yeah were there unique aspects of the architecture to achieve those goals or is it a matter of optimizing architectures and sort of yeah there was there was some um like we had Hardware widgets for all kinds of stuff like the way Apple UI went together was multiple layers we had our display pipe did really sexy compositing um at a pretty low level the way the GPU CPU the camera processor work together was pretty I think novel at the time um but a lot of was just like really looking at the Target and aiming at it and then Apple by designing her own chip they didn't pay the margin stock so we bought Wafers directly from Samsung and then later tsmc rather than buying them from Qualcomm or Nvidia so those guys bought Wafers from tsmc and then they had to Market them up so Apple could turn the extra cost into way more transistors and and we did and so that actually translated not to better necessarily better margins but maybe but also in terms of the size of Chip you could build that for the cost of the Chip and a Apple phone we had twice the transistors okay as some of our competitors which pretty interesting and then the the software applications and the processor kind of co-evolved together and so we made op we made optimization choic that let us do stuff that if you said hey everybody do you want this great camera processor and one person says yes and one person says no Then you put a an average camera processor in there because you can't monetize it whereas Apple was like we we we knew what we wanted all the time and and it was it was dedicated design which is right so you could really optimize the design you had more transistors to work with yeah what about tools design methodology or was there anything unique that Apple had that wasn't available this is one of those things where new teams often do things that old teams can't do so as the apple design team grew we had people from all over the place you know Invidia Intel AMD small companies and then at some point when we were working on methodology stuff it was sort of instead of iterating on the design methodology that everybody been working on for five years it was like what are the best features from all the different methodologies anybody had worked on and at one point I thought we probably had the best design methodology like in the industry and that's because it had people from all different places with different ideas and it was like let's say h hybrid Vigor which was pretty good was Steve Jobs a constant presence in your activity or is he no he's a looming presence L you know it was uh I I only talked to him a couple times like everybody knew what Steve wanted like the next day it was amazing my boss talked to him all the time who was your boss Mike hbert he was s sort of The Unofficial CTO of Apple like we we thought of him as that way but yeah and then we also work for Bob so he was the technical guy in Bob Mansfield was the senior vice president of Mac engineering then later phone engineering then chip engineering and he was a great senior vice president like the best I've ever worked with so and it was really interesting to watch those guys translate what Steve wanted and they talked to him a lot and to get what he wanted and and it was yeah it was pretty much a vision driven company with let's say lots of Direction but not like like Steve's not a micromanager not a how to do it or whatever but just yeah he's not a how to do it he's here's what I want and what can you do and and if you know tell me what you can do and then we'll see what we make out of that so it's a yeah it was a really wild experience so you left why did you leave uh Steve Jobs passed away and then Mike Mike also passed away he was yeah it sort of like I knew it was going to be a way different place and I I partly went there to learn stuff like app's an amazing company and I remember after being there after about a year I thought I still don't really know how to get anything done because it wasn't the same Bottoms Up engineering culture that I was used to from digital equipment and and when I was at AMD and then the startups with Dan they were all the same culture companies and and apple was much more of I want to do that and we're going to do everything to do it as opposed to we're great Engineers you know how to build computers we'll tell you what you get and we'll tell you what you yeah that's and and I wanted to work a place that created stuff out of nothing by intention and uh it was a wild place to work and and uh that was great but I I I thought I learned a lot and I kind of wanted to try it out so I went to AMD which was a a failing company so who who brought you back to AMD uh Rory Reed he was the CEO at the time yeah yeah so I knew Mark papermaster I had met Mark during the power PC negotiations at pme and then U we overlap for probably a year or so when he he kind of came in to run the phone group and then Steve fired him over the antenna gate stuff but Mark was a really good guy and so when I was looking around I called Mark and then I talked to Rory a bunch and I quite like Rory and then I worked for Mark and Rory said I do the business you do the technology I'll make sure we don't run out of money I need a chip and and the goal at that in or AMD was not doing well at the time was doing poorly yes and so it was design a chip was it designed from scratch I mean yeah well I soan nothing R show me this graph so AMD server you know starting with opteron went from zero to 35% market share and rate back to like 4% and he said Jim what happened I said well the product sucks and so nobody's buying it he goes goes it looks like nobody even cares and uh it turns out bulldozer wasn't a good architecture and then it was still being built bulldozer was the architecture for opteron yeah yeah well no bulldozer came later so so the k6 guys came up with this great cell-based methodology which we built K7 and then K8 on which was opteron and then they wanted to take a step further to raise frequency and do the next level and they basically brought back in a lot of custom design and let's say some questionable architectural features and those two things they didn't achieve their frequency goals and then the architecture features were not great so and that created a product that wasn't competitive and then AMD also was struggling with Foundry and they ultimately spun it off to become Global foundaries so then they were saddled with a a tier 2 Foundry and a tier 2 architecture and so I was brought in to be the architect of Zen and then to rebuild that was it clear what needed to be done oh yeah yeah it was super clear the methodology was bad the team organization was bad the architecture was bad and then that that the team would tell me stuff like Jim we can't compete with Intel because our process is behind and I would say well actually the IPC is literally half of Intel's and I'll tell you what if we match their IPC I'll get us a process that matches their process but you can't compare complain about the process if the IPC is no good IP is the uh instructions per clock for clock and so so so there's an architectural component and a process component and like at some level process floats all boats you know so you know IPC is like how good your architecture is and then architecture times process you know performance gives you your product performance and they were missing on both now I had no idea how to get um the process that was good so Mark Mark was a key player on that they negotiated a cross license from Samsung 14 the global foundaries which worked and then later on the Rory's contract said something like AMD will build at Global foundaries as long as Go Global foundaries has a competitive Leading Edge process and then Global foundaries announced that we're not going to do their 7 nmet process yeah they steep back from that didn't they that that set AMD free to go to tsmc but yeah it was a pretty big lift and so I I called Rory up and said said do you mind if I cancel all our current projects so this was the thing that set Zen free so they had a boulder road map and then a Jaguar road map they had a big core and a small core and neither of them were competitive the Jaguar was actually pretty good but people weren't buying that that that zone of processors and bulldozer was not competitive for for a number of reasons and Rory said Jim nobody cares if we're 50% or 53% of the competitor you know they're just buying us for pity so so he says basically do whatever you want and he would call me up how long is it going to take I said four years he goes I don't have that long and uh the board timed out on him and he left but um yeah it was uh it was it was super fun and it was partly an architecture thing like architecturally you know big Strokes there is a spreadsheet that says well you need this many instructions for clock this kind of branch predictor just kind of right you need this frequency we set this bounding box which the team was endlessly trying to renegotiate I said 10 square millimeters 5 Watts three and a half three and a half gigahertz and they would come back and say well we can do 10 millimeters but it's only going to be 3 gz and I was like no way and and then I it's the mission was to turn the gaps in the problems to be solved as opposed to things we can't do that we renegotiate and that mostly worked out and then we had a arm processor that was like architectural companion um because there was a belief that arm servers would be big and I think that that could have been true but it was a little early for arm in the server space and and then we delivered a couple other products we did uh Xbox chip uh Sony game console chip so how long did it take you said four years was too long how long did you yeah it took four years four years and L Su came in as CEO somewhere along the way yeah she she was uh when I was there she ran like the one of the business units and it was armed servers embedded in semic custom and so for most of the time I was there she was doing that she was she was not part of then so at what point did you choose to leave what was you got that project done or will on its way it was mostly done and then let's say there was some management difficulties yeah I didn't really want to leave but you what I didn't really want to leave but like not everything goes your way I'm kind of a bristly character so and you know I like things done well yeah it's hard to say I was going through a phase of like I I knew how to build everything and it was pretty clear so so then on to Intel uh Tesla Tesla yeah so dougfield um work worked at Apple in the Mac group and he went to Tesla and I was talking to him and then Raj Singh and another guy that worked for Doug and and they said you should come here and I was like what do I do you guys build cars he goes no no cars are going to be like a Computing platform there's going to be computers and everything and I was like that's hilarious and then I I started researching like Tesla had an autonomous driving plan and you know obviously um you know Elon was was already somewhat famous even way back then and then I uh so I went and talked to Doug and a bunch of people and it was like come in here you know you can work on electronics so we needed to upgrade the autopilot hardware and do the stuff and I thought we could literally build a chip for this like the chips they were using weren't weren't very good and they weren't you know they weren't hitting the target let's say and then they organized a interview with Elon so I flew down to LA to meet him at SpaceX and I walked by a conference room it was uh lat conference room so my uh my father was the chief engineer on L side years ago which was a satellite so and then there was a Rocket taking off and I'd seen a Saturn five launch when I was a kid and uh so so he like and I reminisced about the Saturn 5 rocket landet and and then I you know we had a pretty good chat and I told him I could build a computer chip and drive car in 18 months and he basically said okay and how long 18 months 18 months and we did it was like a world record is that for you were using tsmc uh the chip was at Samsung so so I took the job I didn't tell anybody there was still some dust settling from leaving AMD and then my friend Pete banon called me up like the day before Christmas or after Christmas and said hey I heard you're going to Tesla and he said how'd you find that out and I said your wife put it in the Christmas letter that you sent all your friends so I said I start like in you know January 5th or something and he said well I want to talk to you about it I said great you want to come over then he says how about tonight and he came over we talked for like two hours and he was still at Apple and he was frustrated with their AI Direction and and so I joined and like three days later Pete interviewed with Elon and and then Elon said he's great and then I hired Keith Whit who was a lawyer cuz they're going to build a chip in 18 months I need to execute like you know 35 we ultimately did 35 IP contracts Supply agreement all kinds of stuff and I called my friends at Samsung I'd work with at Apple and so I'm going to build this chip I need to tape out in one year and they're like that's very hard so It ultimately was 14 months to tape out four months in the Fab two months to bring up the software and we drove a car and Pete had this unb un believably cool idea about the AI engine so at the time the the one of the popular Frameworks was called Cafe and it spit out a list of you know op data data data and then people are taking that and run building a compiler to write a vectorizing compiler on it and Pete said what if we just execute the op and the data as written and we made a chip that basically executed the cafe instructions as instructions and the data might be uh you know 3x3 convolutional data source or something which has a whole bunch of interesting ways to walk through it and we built the memory subsystem to Swizzle the data around to match what Cafe thought it was executing which made the compiler really simple so everybody in the world is hiring hundreds of people to write Ai compilers and Pete wrote the compiler himself because the translation from Cafe output to instruction level output was let's say relatively simple there was ultimately more detail on it and then that gave us a architecture definition and then you I heard some friends from a variety of places to go build the engine and then Samsung built the chip what was the uh sorry um what was the key to getting it done in 14 months getting the IP contract done and just every step of the way every step of the way it was like parallel execution like by then I'd gotten really good at you know I I tell people I'm 100% person so you had a you know there's a AI engine a compiler that goes with it there's IP contracts we had a really good SSC architect David Glasgow Dan Bailey joined me from AMD where he's one of the best silicon guys on the planet the execution team and Samsung I knew those guys we spun them up pretty fast we made decisions really fast like security architecture safety architecture bus architecture we screwed around for a little while on camera processing like AI images at the time were all processed by camera processors but it wasn't obvious that was a good answer um we were unsure about some software stack components we had to comport a bunch of stuff so we used the arm GPU so does Tesla still use the same architecture in the so Hardware 3 stopped shipping something like nine months ago they shipped it for almost 5 years it was the best autopilot ship on the planet for five years and then now shipping hardware for which is upgrade there's architectural similarities uh like we started that project when I was a Tesla but I I'm not privy to the details of how it came out so did uh did you have a lot of interaction with Elon along the way or not at all or yeah yeah he was there two days a week and we talked to him all the time and then I had autopilot hardware and at some point Doug asked me to run low buge electronics and then on and off the autopilot software people reported to me depending on who'd been hired to let go recently so that was you know let's say dramatic and then did Elon want to insert his own ideas about how things ought to be done or understand the architecture and yeah so at some level he wasn't a chip guy he had written software he's really good at physics mechanical engineering and visualizing things and it's a lot of the chip stuff I'd explain it to him he'd be very interested but he didn't have lots of ideas about that but how the overall thing worked and why yeah he was he was really interested and it was interesting to explain it to him MH I learned uh so a lot of places you know you start with here's the problem statement and here's the data and you kind of walk people through your journey elon's a solution first guy we're going to do this that's cuz we had this problem here's back everything else is back up and one day we we're walking them through here's the here's the images we captured at night here's what's going on here's why they're hard to Pro he just lost his mind like what are you guys talking about well we're going to get to the solution well what's the solution it's on page 18 well start on page 18 like what's wrong with you it's like here's the old picture here's the new one he goes That's great it's a way the new one's way better why is it better oh let's just go through the presentation backwards and uh but it was now is his directness was great like when he talked about about first principles I've said this before it's like you think you're a first principal person you talk to youone it's like no no he meant Adams you know like like first principles is like a whole another level right so what got you out of there uh we finished Hardware 3 um started Dojo but it was clear to me it was going to be a while before so elon's a big step guy and uh i''s say chip Engineers you there's a road map Cadence and continuous Improvement and I thought I wasn't quite sure I was going to do for two three years before we taped out so so Hardware 3 to Hardware 4 you can read about in the papers was was what you can read about in the papers it was like four to five years apart okay like as far as I was concerned we're going to tape out Hardware 4 a year after Hardware 3 and you know make you know steady improvements but like elon's more of a take a big step use the out of it take a big step and there's something to be said for that but I thought I want some new excitement and then I got offered SVP at Intel to run their silicon Engineering Group and and I thought that would be wild because Intel had the best Fab the best CPU technology the best they had Cross Point memory they had a whole bunch of stuff and I thought we'd go do the next generation of high-end servers and uh not any none of that happened but we uh it was a really fun job so what made it fun fun well when I left AMD the team was 2500 people Tesla was you know four or 500 people I joined AMD my team was 10,000 and join Intel yeah I joined yeah intel was 10,000 and the scope of it was epic you know client parts and server parts and networking parts and and then I totly figured out that a lot of in technology was great but a lot of it was stuck on old cat TOs and old design flows and yeah Intel really got in a rut with their design tools so so I was like well how do we re-engineer an an organization this big to be oriented properly and they had done some stuff like you know they used to make integrated client part so the CPU team the memory controller team PC Express team they all work for the same guy they all work together and they made one part and then uh Mery who was EVP said we need to have IP teams and S so team and CPU teams and he broke them up but the way it it never really got refactored like the IP team wasn't really IP the IP was just a Design Group that delivered stuff to thec who tested it and so they they had all these things like half transitioned and the challenge was like how do you re-engineer that and fix the cat tools like I got a call for art GS he said the Israeli team is using our new cat tools how' you do that I said a lot lot of fighting and uh but it was fun like those guys were great they they built literally the best design flow on the planet and used it for 25 years and you know it was very difficult to say the the world's moved on and and change is going to come and like how's that happen and and some people were like Hey we're ready for it and some people obviously it was uh you know people were very emotional like Engineers spend their lives building stuff Intel still had a big internal CAD group right yeah 1,200 people and a lot of them were just building wrappers around tools and so we introduced this thing we called the Bare Metal Flow which is like how do you build low-level interfaces for all the cat tools and like how do we change how we think about C tools and you know some of that worked out pretty good tense torrent yeah so tense torrent is an AI company so how did you did you burn out at Intel with 10,000 people or what was the no it's like at a high level I I thought the company and the engineering team needed a let's say a much bigger reset than the CEO did and I I quite like Bob's one we're we're colleagues but like his vision of what needed to be done and my vision were pretty far apart I wasn't burnout I was but you know you couldn't get done what you wanted to get it was very difficult and I like i' learned you know learned you need to be in align with you know the CEO runs a company and you can't be like going to a meeting and say we need to do this and they're like the CEO said the opposite and you know he had his reasons you know he was financially managing the company and I was looking at a whole bunch of key engineering technologies that needed a pretty big overhaul that was more than you could do say just on the way like at AMD with like Rory and I was like we need to overhaul this from scratch and he said great go do it because what's what's happening is not working so like like that worked pretty good and until it wasn't working so so I I thought I wanted to go work on a startup I thought of starting a company I had been the angel investor at T torant I like the AI technology at Tesla I got to see every is that Tesla was your first introduction to really be inside AI technology yeah even there I wasn't the architect of the i engine P was you know I architected the plan to build an autopilot ship and you know I'm I'd say very good at working across functional boundaries to make things come together and I thought I could help T torent launch the product and also build the Next Generation and then also T torent AI Technologies built on RIS fire processors and then we use multiple RI processors to to enable that from very small ones to medium ones to big ones so I also we raised the investment to build a high-end CPU team to build a high-end like basic basically 10 pole rist 5 processor to basically what the 10 pole so rist five started in Berkeley amazing quality the boom core and the rocket core or open source RIS 5 processors then SciFi Andes vantana built I would say more product oriented mid-range processors but nobody built a eight wide out of order high-end RIS five processor and the CPU world that the people want to go from very low to high The Genius of arms IP plan is like a lot of CPU companies every time they build a better one they drop the old one and arm is like they got M Class R Class A Class you know server class and so RIS 5 needed to extend that line so I thought it would be really interesting to bring build A5 based on RIS 5 you know computers at the bottom and but also build the co-processor and RIS five to to put that together this all has to be done in the context of how you deal with uh Ever Changing AI models and algorithms right yeah so so it turns out the AI stack is really complicated from top to bottom it's not like a person has an idea and they write a c program that runs through a compiler to a chip there's there's about four or five levels of translation like pytorch programs run very slowly like a 100 KZ so like it takes you a while to get your brain wrapped around that so you're going to put a p flop a computer on a computer program that's running at 100 KZ cuz it's parsing Ops and the Ops are really big and then there's translation from the pytorch level what the programmer thinks he's doing down through the software stock and then you need to build that in a way it's flexible and resilient to the new ideas they have at the high level but translates into high performance Computing at the low level and that's the that's say intellectual AI challenge that t stor you know working on so did tin t start to build a processor to build you know what was build an AI engine to build an AI engine for what application for general purpose AI for computing like the basic data center level or yeah yeah the basic engine is a tensor processor there's like generically three kinds of computers scalar vector and tensor so ai ai is tensor operations on tensor which are multi-dimensional arrays of data and the core of it is a tensor processor fed by RIS 5 processors that interpret let's say one layer of the software stack and feed the operations to that processor and then move data between processors so the yeah the our mission is to make really general purpose Computing that's accessible from top to bottom so every layer has a way for a human to be involved or the software can take care of it so how is the Mission direction whatever evolved over the time you've been there first as CTO and now as CEO yeah the original like so startups goes through phases like I realize this is it seems obvious there's a research development phase here's the big idea well there's the crackpot phase which is I have this big idea I have to raise money but then you go do research and development to build something that's worth doing and then you transition to engineering like make every piece of that really good and then you transition to to product I have a product that somebody could buy and then you work towards being customer focused and then ultimately big semiconductor companies are production Focus let's say so I joined I I guess to help transition from research development phase which the founding team was good at to let's say engineering Excellence phase like end to end and this year we're starting the ship product to be a product company and kind of move through that journey of each step properly but who the customer is has evolved over time right it's interesting the software generally speaking at the at the AI application Level has gotten a little simpler um the original Inception reset were pretty complicated graphs of computing the modern like Falcon llama mstr are simpler Transformer models they're still sophisticated and the oppsite they do um they've mostly grown larger way bigger training sets bigger uh training times uh people want to be able to expand quickly from a single chip to many chips you know depending on the application um I'd say it's still it's still changing at a high level especially compared to the traditional compute server client world but it's changing in me in directions that are let's say understandable um and you've described the company as a design company yes yeah we design AI processor AI software and CPU and I want to be really good at all aspects of that top to bottom we design chips we Design Systems like we we are kind of an end to end company like one or two of the startups have turned into like we're going to put our AI behind the cloud or like semiconductors are hard but now we're a design company like we are friendly with software design tip design architecture design system design and like I want to be really good at all of it do you have a sales force yeah who are they calling on well it's all over the place so so the mistake all the AI startups they say we're going to build this chip we're going to go sell to you know AWS right that has to run everything on day one it turns out there's literally thousands of people who buy AI they put it in equipment they put it in small data centers they put it in research Labs people want to buy AI put it in their own chips like people there's the diversity of customers is really high now the Wall Street guys focus on the 100 billion of high-end Computing for data centers but there's literally billions of dollars of business all over the place and usually new entrance you can bang on the door of the Cathedral of the incumbent or you can go find new customers that are doing new different things and the big guys can't focus on the little guys like nid is not looking for another $1 million customer right but but I sure am so the million dollar customer can't afford a million dollar system so you've got to have tiered systems right yeah you have to have different price points and also you know our sales guys would you know you get 10 $1 million customers is 10 million sales it's real like i' be i' be thrilled about that and I've done I've done this before like and you also can't pick and choose the winners like when we design parts for side bite we had a hundred design ones at one point and some of them became quite big for broadcom and it wasn't obvious which ones were which so like I I think there's going to be another generation of AI going all over the place and you know I want to be one of the companies that has AI that works software that we open source our software stack and it turned out that that's a great recruiting tool cuz a whole bunch of people they want to write software and publish it and they review our code before we hire them it's it's really kind of a wild experience and because we're not the incumbent we're not proprietary we're open you know we have access to all kinds of people that you know incumbents don't have access to could Apple be a customer at some point yeah maybe but they have 5,000 people working on chips and they're experts and I'm sure they're doing their own thing other processor vendors it's possible yeah I I've been through multiple transitions like this when we built optron and K8 you know nobody wanted it everybody knew servers have backplanes right we built a computer that didn't have a back plane know 20 years it's been the dominant server architecture so do you if somebody wanted to buy something today from you what could they buy what could you yeah we're we're selling we're selling our grce skull chips and we're about to announce our our next chip that'ss Network together and we'll be able to hook up you know like like a single chip to 100 chips together and so your gr SCH chip is a single single chip PCR Express device and then Wormhole has 1600 gbit ethernet ports and we hook those together and at the end of this year our third chip uh black hole will be ready to ship and that's like a shrink with a whole bunch of performance optimizations and people can license IP from you as well yep yeah and it's yeah it's been really interesting to talk to people because some people they want to buy the chips to try them out but they really want to build their own product and they want to license IP and obviously Nvidia won't license their IP to you to build products that they would think compete with so you'd like to try it out makes verify performance and everything else in chip form and then as the Next Generation embed that IP in some of their own proprietary product yeah and RIS 5 is doing the same thing is people are using RIS 5 processors for all kinds of little things but we're going to tape a chip out at the end of the year that has our big risf processor so people can try that out and then either use that chip in their product or license the IP to put in their product could that risk processor be used for something other than an AI function or is it of course yeah theis Specialized or optim no it's so I want the risk five technology so like when I build AI I have all the pieces and I can modify the way I want but it's also pretty general purpose that we can do almost anything with it so do you enjoy being CEO or CTO or both yeah I like it's been really interesting to do like CEO is kind of wild because you're sort of responsible for everything yeah sort of and uh and uh and it turns out to be you know the the company scope is way smaller than Intel or you know AMD or Tesla but uh like say the breadth of it is higher like how big are you now yeah 500 people you work in the office or virtually or all the above yeah we have Offices here and there so I'm in the office I'm at home a lot of times the calls start early in the morning so I start doing calls from the kitchen table at 7 o' and then do you care whether people are in the office a little bit um I'm really impressed with how much work people get done with there's you know being wherever they are uh for some people the flexibility is life-changing you know people have lots of needs family taking care of their parents you know they live far away there's all kinds of reasons for doing it I think it's really healthy for everybody to get together in the office periodically they like it like every once in a while the office is full and everybody talks like non-stop like I think something like 2 or 3 days a week is really good for human beings but the flexibility like you know they drop their daughter off or they pick their mother up or they have a doctor appointment or and it's just sort of scheduled in and nobody cares that seems also very healthy so let's spend a few minutes on uh Jim the person I heard somewhere along the way that you had a wife who sent out Christmas cards disclosing proprietary information yes she did it only proprietary to me yeah when you're married you been married for 20 years I have two children they're going to be what 19 and 20 this year so they're off to college one's in college for two years and one starts at cowp in the spring in the fall doing what are the two of them whatever the hell they want to so but yeah they they seem to be both smart and alert and and uh yeah we'll see what happens where did you meet your wife in your busy she was a friend's sister so I I was at a 40th uh 40th year birthday party and I met Bonnie and somehow we kicked it off hit it off so 20 years in huh yeah yeah 20 21 maybe this year somewhere in there yeah that's been great what do you do when you aren't uh hiring raising money inserting yourself in meetings I I've always been into being physically active so looks like you're in pretty good shape I'm in reasonable shape you know I'm starting to get a little older but you know I cut Surf and like snowboard and lift weights and I had to stop running cuz um one of my niece started bothering me and I talked to two 85y old guys and they both told me they wish they'd quit running sooner and I said well when did you quit running the one guy said 80 if I was doing it over again I'd quit at your age you know because you know the KN year wear out and your Replacements hard so but yeah like my wife had her knee replaced and wish she' done it sooner yes mostly it works out good but like I said there no I've been pretty active you know most of my life so so you do a lot of reading on and off you know sometimes I feel like I read all the time you know it used to be like a couple books a week and uh now it's like there's blogs there's the internet there's articles there's this there's infinite sources of stuff and people really pull their well they still pour their hearts out and I tell people to read a lot and uh the funny thing is people always ask me well which book should I read and I think all of them like like and and it's the problem is even if this is a really good book that book in the context of a hundred other books is a much different book than then the book Standalone and you know as a young person maybe you don't read that much but I've I've literally met people in the last month that I thought Christ if they read 10 books their lives would be completely different like there's literally problem solved and and it's not just the way the book I recommend books that I don't agree with all the time but you know the you know the the thinking in the alternative like there's so many things to know and to think about and so yeah I still read quite a bit being that you're still relatively new in the CEO role you've had the opportunity to observe lots of CEOs in the past is there anyone you work with now or that's a good Mentor for you uh you know that anybody rely on for advice or well I work with really different people and you know I've it's really good to work with different people and then learn what you can but but then you have to go do it your way and go figure it out like when I went to AMD I thought well I'm going to try this and this and this and this I actually hired a consultant like CU I was like going to scale from a very small team to a very large team and and some of the experiments were great and some of the you know management books like I had this I asked a few friends AMD like how many management books had you read and the average number was like zero and I'd read 20 I'd read literally 19 more books and the like one person said one I was like wow I literally know more about management from books and it's not like you know books are perfect but you know the top 20 management books in the world you just go to Amazon you start googling for management books and you read the reviews and some of them are amazing and some some some of those books are written by people with 40 years of experience they're passionate about and they poured their heart out into it and you can literally read it and then you try it and it's like amazing how often it just works even if you don't even know why but but you learn from that so so yeah I don't I don't know if I have a favorite I really like the way a whole bunch of different people worked like working with Doug Fields is great I like the way he managed people and how he cared about them and he thought really carefully about it but Bob Mansfield was crazy smart and he always knew when something was wrong he always knew when somebody was making something up like how did he do it and like and he had a very clear style like which was very different from them I got I really like working with Elon like I learned so much from that guy it's amazing but my management style is like way way different I'm a like a person enaer Gap filler and elon's like a force of Nature and uh so I don't know what to learn from that except that I learned a lot and I changed a lot and for Back To Top