Hi, this is Wayne again with a topic “The Future is RISC-Y”.
Look, I usually have to sell it a little, but I am genuinely extremely excited to have you on the show uh. They asked me when I got on the phone with your team, because we reached out when we saw the the dev kit that you guys have right now, so this is on the tense torren side um and I was like okay. Obviously, whatever Jim’s working on is probably cool, as so maybe there’s something here, and so I scheduled a call, it’s just exploratory call and they go oh well, you know, do you want to talk to Jim and I’m like? Well, we really take guests on our show anymore, but um. Yes, we’ll make an exception.
That would be great w. I don’t want to waste anyone’s time, which is actually a big part of the reason we don’t take guests. We are notorious for starting our show anywhere from 1 to three hours late and we hate doing that to important people. So without further Ado, I want to get into some of the community submitted questions. We announced that you were going to be joining us and it would have been a huge disappointment if you weren’t here, but you are so Dylan asks: hey, Jim I’m, a junior Computer Engineering student about to start my first internship doing verification engineering at a big chimp Company, first of all, did I just say: chimp company anywh, who doesn’t matter.
The point is congratulations Dylan. He says it’s great to see how far open source has gone. Uh. We even learn risk five in our introductory Hardware course, oh cool.
So, first of all I want to. I want to start with letting you talk about risk 5 a little bit because that’s obviously a hugely important part of what tends to is doing right now, and I guess I realized. I didn’t really talk about tens torrent at all.
So do you want to give us a short introduction to what exactly Drew you to this company and to their mission? Wow? Okay, so Tor is a AI computer design, company we’re designing a high-end AI engine and also a high-end RIS fire processor and I think yeah, a AI has gone through a lot of evolution and you know he started running on CPUs and then gpus and then I Think Google anoun the tensor processor in 2015 and we’re building essentially array of tensor processors, that’s programmable, with open source, uh software stock that we released in January and then there’s going to be a combination of AI Computing and and general purpose Computing tied together and we Decided to make a high-end RIS five processor, our AI processor, also uses little RIS. Five cores, you know to drive the execution of the big tensor processor so and yeah. So the and the R 5 thing is really interesting because you know at some level computer architectures are generic.
It doesn’t really matter very much if it’s x86 power, PC myips Alpha arm or rist 5, but only rist 5 is open and the Berkeley guys that started it were pretty good and the cool thing about open source you know we saw with Linux – is when it’s Open source, a whole bunch of people can work on it. It’S a it’s a much better Innovation platform and there’s a it’s a one-way door. When people go from, you know proprietary technology to open source, they literally never go back. Linux killed literally all the proprietary Unix operating systems and I think Solly RIS 5 is going to take over the Computing World, which is pretty fun just want to say your student friends go go ahead. I was just going to say if I’m one of those Berkeley guys I’m putting Jim Keller, said, I’m pretty good on my LinkedIn like immediately. That’S awesome, they’re, pretty good, well yeah! I can tell funny stories about, like you know, computer science in universities and computer science and high-end computer design companies. They they kind of work together, and it’s really interesting. Cuz, a team of a 100 people worked together for 5 years, can refine that c out of somebody something whereas students, you know they get a project and sometimes without that much support, and some of those projects are pretty good and some of them it’s hard for It to add up to a lot, let’s say so you, but like like the branch predictors everybody uses came out of universities and RIS. Five architecture, which is going to let’s say, dominate Computing in 10 years years, is came out of universities and now there’s 20 odd companies, building RIS five computers and way more using it. So, whether you did it on purpose uh or whether you did it by accident, you actually transitioned me perfectly into the second half of Dylan’s question and uh.
This is really cool cuz. He asked for getting into a higher level architect SL designer position. If you want to work on one of those teams, are you going to recommend stay in school, go for the PHD, or do you want years in Industry? What are you looking for? I well so. Phds are really good for some people.
If you really have a research topic and you really want to go think really hard, but if you want, like, I didn’t study computer design in college, I’m an electrical engineer. You know my my major is first electromagnetic fields and then, when I, my my advisor, ran the the semiconductor physics lab. So I learned about a lot about that and then I took one logic design course and then got a job doing that, and then I got a job at digital. Where I work for a great architect, Bob Stewart and then computer Architects, good ones, know about a lot of different things, so I learned how to program you know do lot to design. I know about semiconductor of physics, I I know a weird amount about packaging and you know signal integrity and all kind of stuff, and so, if you want to be a computer architect, you should probably work on a lot of different things and most most computer Architects. That are really good at it. Didn’T do it in college.
You know as a PhD interesting enough. It’S it’s almost like a too narrow of a way to go about it. Phd guys tend to be experts in something I Mak computer.
Architects tend to be tend to be generalists, I’d, say yeah, so you already kind of alluded to this one as well, but William asks I mean you’ve obviously got experience on the arm side. You’Ve got experience on the x86 side. You’Ve got experience on the risk.
Five side um William asks: how far do you think x86 64 can go? I mean you’re telling me now you’re saying look. Risk 5 is going to be the future. You gave that number 10 years, I’m not going to hold you to it.
I mean I can’t promise nobody else will, but I’m not going to hold you to it. Um is that because x86 is out of gas, or is it because risk 5 has just got some kind of fuel that we’re only just discovering the potential for in the engine which, which one is it neither so computers generically? You know they fetch instructions, decode them and issue them right, and the thing that makes the front end of a computer fast is how many instructions can you decode and how well can you predict the instruction stream right so x86 has a deficit in the sense that You know random links. Instructions are harder to predict, but we sort of figure figured that out. It’S just harder to do, but it’s not like a big limitation and then the execution engine goes fast, because you have lots of parallel execution units in out of board issue which is generic to computers.
And then you have a good memory system with a really good predictor, for where the data is coming from, which has nothing to do with your architecture. So I’d say: x86 has a limitation, so it’s 16 registers variable length, instru set and it sort of has a pile of old crud that nobody actually needs, but you have to build so so it has a tax right, but computer performance is mostly today based on Prediction and the number of predictors in a modern computer is crazy. We predict obvious like where the instruction streams coming from where the next branch is the direction of the branch call return stack. We predict the width and grouping of instructions.
We sometimes predict the results of instructions. So tell me this where the data is coming from. So it’s all a prediction: uh one of the uh, a follow-up question from William is you know, could we see cores using multiple architectures? Could you see someone like an Intel or an AMD, an x86 licens holder taking uh uh, some kind of risk, 5 architecture, AI processor, co-processor, and using that for prediction like an AI accelerant on a traditional? Does it just not make any sense? I mean tell me if, if it’s a dumb question, I’d love to know yeah, it probably doesn’t make very much sense. So computers are very optimized to run.
You know a particular instruction set today, there’s pretty good binary translators and they keep getting better and binary translation from like an arm instruction set to wrist. Five is relatively easy and back and forth so you’ll probably pick your general purpose Computing architecture and then either recompile. All the code or translate the code you need, like apple, switched from x86 architecture to arm architecture. They hardly missed the beat nobody even noticed or cared that’s the ecosystem and sofa build, but they could switch to RIS five and nobody would notice or care either.
This is hilarious uh. He keeps beating me to what my next thing is going to be. It’S all prediction right, so I just don’t. Even why do I even? Why do I bother talking? Well, we live in a simulation and a good simulator predicts everything, and so so the next thing I was GNA the next thing from William’s question that I was going to focus on so so he asked okay.
How far will x86 go? Do you think armor risk 5 will replace it in the future? Uh, maybe we’ll see using multiple architectures. So that’s all Williams question so far, I’m not taking any credit for that or blame, but what I will take credit or blame for is this next one? I was really focused on the word replace in his question and you brought up needing to recompile code, and software is something that I feel like is a a bit of an elephant in the room. You know when you talk about how well you know. Fundamentally, all processors are the same.
Essentially, it’s like how many instructions can you process, but while Rosetta 2 was an absolute Marvel um to the point where, just a few years ago, I wouldn’t have even shortlisted Williams question. Let alone asked it to you uh, but now that I’ve seen what Apple was able to do with that x86 to arm transition um and what Qualcomm is claiming that they’re doing on their upcoming Snapdragon chips with Windows on arm um, I feel like anything’s possible and That word, that replace word can I expect to go back with with to Legacy programs right um to to stuff that that is that that tax on x86, that taxs on Windows and whether it’s through uh, whether it’s through AI or whether it’s through on the-fly recompiling, Can I expect to replace the gaming PC that I have today with something risk 5 that will run and I’m not going to ask for 100 %, but if I asked for 90 % of the software I used to run, do you foresee that yeah, of course, Of course, more and more, more and more software is written in more and higher level languages like recompiling C programs in Java and py py Python, and you name it it’s getting easier and easier, like the architecture mostly doesn’t matter now what matters is on a a Given architecture like we found this, we started building like a server stack for RIS 5 and when they went from Intel to AMD to arm to rist five each time you port software, it gets easier to do and the hardest Port by the way was Intel to Amd, even though they’re both x86 right really and that’s because there’s a whole bunch of proprietary software in the server stack that was actually Intel proprietary. So you weren’t, which, by the way they weren’t giving out the poort. So they had to rewrite a bunch of stuff. But all the new software is C C. It’S clean, so porting arm was easier right, porting. The risk five is pretty easy um. The thing you find is like the tool chain.
Maturity like somebody built a binary with some set of switches, and then you link that and somebody mislabeled one of the header files, and then you have to be an expert to figure out why this thing didn’t work. But the actual porting of the software is not not the hard part 5 ecosystem uses GCC and lovm and they’re really mature compilers, like they literally use the same compilers. On the back end for all the architectures you mentioned there, the server stack for risk 5 um. That’S a huge deal, and I know there was there was the struggles with the Intel, the AMD transition and that hampered AMD a bunch of stuff it it’s going to be a big problem. Solving the the server stack thing how’s that going. I know you guys are working on it. I know some other companies are working on it, yeah, it’s going pretty good.
So and again, this is one of those. So so Amazon did a really fun thing, so in AWS they put graviton in there yeah and first they they ported some of their own applications so and Amazon’s pretty good at putting a gun to somebody’s head and say you will go. You know Port the software and get it running, and then they said yes, sir, and they did you know and but then they put it up on the web and said hey.
If you want to you know, P your stuff to arm it’s 30 % cheaper, whatever the number was and a lot of people said sure, that’s easy. It’S JavaScript anyway, who cares right and so people started porting it and the more people ported the better it got, and it’s easy to tell if the application and they made it pretty generic. So I think what you’ll see is like heterogeneous data system so you’ll. So you have a cloud and there’ll be some Intel servers for the dinosaur code and then there’ll be arm and wrist, five or stuff. That’S already been ported and there’ll be a price difference, and then people will go. You know where they need to keep converting. Nobody cares about IBM 360 code or vax code or sun code, or you know HBU code like it’s. It’S all gone and and and you won’t care about the games that you ran 10 years ago, because there’ll be better games, they’ll just emulate them or they’ll AI emulate them like that’ll, be the really funny thing you’ll say hey.
I want to go play Super Mario Brothers and you’ll talk to a computer and describe Super Mario Brothers and and play a little. Video of of of you know, YouTube video from the 0s. You know that play Super Mario Brothers and it’ll emulate the whole thing and and you’ll think it’s fine, that’s kind of terrifying uh. I can tell you, there’s a there’s a lot of Gamers watching this right now that are going to be really unhappy about that, but we’re going to move on we’re going to move on guys. Some of them could be really happy because now they have, these games are going to get pong from this. You know the 80s and and they’ll be able to like train an AI and to play perfectly Nintendo’s litigating, already Space Invaders, all the stuff that used to play.
I I can hear Nintendo’s lawyers from here yeah .