Tác giả giữ toàn bộ bản quyền (All Rights Reserved) — không phải giấy phép mở. Bạn vẫn xem và học miễn phí trên OpenER; muốn tái sử dụng cần xin phép tác giả.
computer and start webinar. Here we go. And and get back to the first slide.
Quick, get back to the first slide, click, get back to the first slide. Okay.
Welcome students. Hello, hello, hi.
Everybody. All right, what a special day we have, what a special guest we have today. I'm so excited about this. Let's get started. All right, I think we have the numbers are 40 people, which is about right. We usually have 30 to 40, about 50. Wonderful, welcome everybody to CS 61C week 14, Warehouse Scale Computers. Wow, great to have you all
here, great to have you all here. Uh, let's share, let's jump right in and talk about the agenda. So uh our amazing guest is Professor David Patterson. I believe this is our first Turing Award winner, Bora, on our on our as our guest panel. We've had some, we have the DARPA lead, we've had CTO of big companies. It's been an
amazing semester, um, but this is the first time we have a Turing Award winner, which is the highest, it's Nobel Prize in Computing if you don't know. Uh, so we're going to jump into David in about 10 seconds, but the agenda for today is after talking today we're going to have a Computing in the News as we always have, then midterm score update, we have some midterm scores to release to you and we have some some news, so that's exciting.
Double-do score a schedule and then ask us anything.
So boy, it gives both Bora and myself and everybody on the staff the greatest of pleasure to invite Dave Patterson, our colleague and friend who helped build this class. By the way, this isn't just a random Turing Award winner, this is a Turing Award winner who has contributed significantly to, Dave, I think you also won the Karl award, I think you won like every award ACM has to give. Maybe one
of the most honored people in the entire department, arguably the whole campus. Um Dave retired recently, Professor Emeritus now, and he's now working as a distinguished engineer with Google, some exciting things, maybe he'll share some of the fun things he's working on with there. He's the vice chair of the RISC-V Foundation, um, and he's known for partially co-inventing RISC,
RAID, 61C, the textbooks that are the most popular textbooks in the entire planet, by the way, in computer architecture. There's no textbook that's more popular than Dave and and Hennessy's, uh, book in architecture. So what a gift of his time, uh, to, and by the way, Dave was the first person I taught with in 61C. He was the one who taught me what 61C was like, so I owe him
everything I am in 61C. I owe to Dave. So Dave, thank you so much for the time and for joining us today. What a delight. Oh, thanks, it's always great to try and catch Dan's enthusiasm, because, which is contagious. Uh, I thought I'd start off about, talk about five minutes, tell you about RISC-V and then we're just gonna open up for questions. Sure. Oh, I can see
the slides. Cool. Yeah. Oh, perfect, perfect.
And by the way, you have more than that, we have a lot of time. We have, I think I'm, I'm looking at about 30 minutes, so don't feel like you have to rush through this. Don't rush through it. I'd be happy to do questions and answers either from you guys or from, wonderful. So, uh, the story goes back to 2010 and there's a research project right there in Soda Hall where it was time for
us to do computer architecture research.
In fact, Borivoje was involved with this at that time, and this was led by our colleague Krste Asanović, and he wanted to do computer architecture research, and as you know, you know the vocabulary of computers is the instruction set architecture. Well, the two obvious choices were the x86 and ARM, but it was
impossible for two reasons. First, they were very complicated commercial instruction sets, thousands of instructions, and even if we wanted to, uh, they protect those instruction sets because they own those instruction sets and they don't want to let anybody have it, even us. Even though Intel was one of our sponsors, they wouldn't let us do
that, so we had to develop our own instruction set. So Krste Asanović and two grad students, Yunsup Lee and Andrew Waterman, decided, okay, we're just going to slap something together, three months. Well, it took them four years, although along the way they had to, they built some hardware. So it's, it's an instruction set that's informed by actually trying to build it, and it
actually got better over time. It actually got simpler over time, surprisingly. And the original idea was, well, as researchers we want to use it, and lots of other researchers will have our problems, so we'll make the instruction set, which as you understand is just a specification, right? The instruction set tells the hardware how it's supposed to work, but then the hardware itself is a separate thing that
implements that instruction set, like Ethernet and, Ethernet standard Ethernet controllers. But, uh, there wasn't a lot of uptake with other academic units. The surprise was, four years later, industry started showing interest. They really liked the idea of an open instruction set, and once we started talking to them to understand why this industry was interested in our kind of research
instruction set, then thought, well, that's a great idea to try and have an open instruction set. And as an analogy, it's just like an open source operating system. Linux is probably the most popular operating system today. Could we do something similar to that for an instruction set architecture? So Krste and I wrote a
technical report, uh, "Instruction Sets Should Be Free: The Case for RISC-V" and I've learned recently, uh, because we just celebrated the 10th anniversary and talked to a lot of people at the time, actually a bunch of people read that paper, and that got people to decide, okay, we're going to try, we're going to use use RISC-V because of the arguments in that paper. Okay, next slide. So let's go over RISC-V, uh, uh,
you've been using this in the class, right, Dan? Y y y yeah, yeah, okay. So what do I think is really important? First is free and open, anyone can use it. I think that means since everybody, many people are using the same instruction set around the world, uh, it's more competition and so I think it'll be more innovative, particularly for startup
companies or small companies. For proprietary instruction sets, the first thing you have to do is sign a contract. They say, well, okay, I'll pay you this much money up front, and then we'll give you a fraction of every chip that we ever sell, so you want to negotiate that contract carefully to, to, you want to be profitable, so that takes a long time and it t- besides money. So
what's nice about RISC-V, if you're a small company or a startup company, you can just say I'm going to use RISC-V and I'll figure out who I'm going to get the processor from later, because anyone can use it. So that's a kind of a big business advantage independent of its technical qualities. Secondly, I think of it as it's really elegant because of classes like 61C, exactly, 61C. When you
try and teach assembly language programming, you don't want to show people the whole instruction set because it's, you know, can be hundreds or thousands of instructions and there's no real pedagogic value to that. So what you do is you subset it, you say here is the the essence of that instruction set. RISC-V because it came came out of academia for teaching as well as research, that subset is official, so
those 40 instructions in RV32I is the whole is the whole simplest instruction set. So all RISC-V software will execute in just those 47 instructions if you want.
So that makes it, it's a new idea for instruction set architecture. Usually what happens is they start small and over time they get bigger and bigger and bigger, kind of as you'll see as, as you get into middle age, your bodies will probably get bigger.
But the instruction set RISC-V is going to stay skinny for its whole life.
How does it accommodate other applications? It has optional modular instructions, like you've seen the floating point instructions, or the atomic instructions, or the compact instructions. Those are optional if it matches your situation, but you don't have to use them. So that's a new idea for instructions that, in the past, every time you added a set of instructions, every implementation of
that processor for the next several decades had to also implement it. For RISC-V it's optional. So those are two: a simple core and modular options are two, uh, big important changes, because that can make it much more targeted for the application. The third big idea, which is kind of easy to do in 2010 when we were
doing it, is we knew it was going to be important to add accelerators to instruction sets. We knew Moore's law was slowing down, we knew that this change was happening in the technology, so you're going to need accelerators. So basically, uh, we just left lots of opcode space in it. Now it's, that was easy to do now, but 20 years before that people
didn't think that was important, so they kind of filled up the opcode space, so it's hard to add more instructions. It's easy and RISC-V. Not surprisingly, uh, in 2010 we thought the future was at the edge and in the cloud, so we wanted it to work well in the cloud and at the edge, so it has both a 32-bit and a 64-bit address
shockingly, because the biggest mistake you can make in computer architecture is run out of address bits. That's not going to happen. RISC-V, because we also have a 128-bit address space version. We're not probably going to need that for a few years, but it's nice if, because if instruction sets are successful, they can last for 30, 40, 50 years. We've already,
when it's time to go to 128-bit addresses, we'll be there. And finally, because it came out of academic roots, and once the graduate students graduate who worked on it and faculty get distracted, then it could kind of fall apart because the faculty aren't paying attention anymore to that project. Like Linux, we started a
foundation. It's called RISC-V International, and it owns that instruction set and is evolving that instruction set. So that's why it can last for decades, even though academics can't, you know, keep that much attention. And then finally, in this kind of new world, people are worried what's in their computer hardware. They're worried that there's some trap door in
there doing something nefarious they don't know about. So by being an open instruction set, several countries all over the world have had those concerns.
They're worried about getting technology from a competing country that might, you know, be an adversary. So you can design your own core, and also because the open architecture allows open source implementations, like written in Verilog or SystemVerilog or something like that, you can inspect it, you can know what's inside your computer. So there
aren't any secrets. So these seven bullets are what distinguish RISC-V, and while there's a lot of enthusiasm for it, okay, next slide. This is RISC-V International.
It has more than 600 members in 50 countries. It's a really worldwide phenomenon. Depending on who you talk to, they think, oh, this is our continent. You talk to the Europeans, they say, oh no, it's this, RISC-V is primarily European or in North America, well, it's primarily North America, Asia, oh no, it's we're the we're the center, right? So they argue about it and there's 2,000 people participating
in the various uh projects to expand the architecture in different modular ways.
So to try and convey this enthusiasm, the next slide we'll see if that that is a video. Uh, it's the 10th anniversary, we were going to have a party that quarantine canceled, so we put a video together instead. If you go to this website, you can, there's a variety of videos that we've done kind of to document history. This is the three minute version that tries to capture the
enthusiasm. So if you guys could play that, that'd be great. Yeah, Krste is gonna take over and do that here, see if it comes through. All right, song beautiful, it louder, what
is mach, that's as loud as it goes. Okay, that is so energizing that for those who haven't joined the community yet, come on in, it's a really, really powerful place to be. I'm actually quite impressed the kind of energy that is going, so many
parties, the very lowest level have a common lingua franca. If there is a lingua franca, then we can start building standard computers. Computers are getting, oops, sound, uh oh, uh oh, uh oh, I I clicked the mute by accident, I'm
sorry.
Me, sorry, my bad. You can't build a society on proprietary formations. The key differentiator for RISC-V from other processors is the openness, the fact that you don't need to sign any agreement whatsoever to just start using it. With this you can just
download it, you don't have to ask anyone's permission, and that is a fantastic difference. Now that I'm doing architecture, use RISC-V and I get so many things for free, I can focus on adding value on a few instructions and leverage tens of millions, hundreds of millions of dollars invested over the last 10 years. You go
that much further by being able to leverage that infrastructure. We had a workshop where we announced that we are going to RISC-V in every GPU, every chip that we have starting this year. 95% of resources are put in RISC-V development.
When Western Digital announced that we were going to transition a billion cores, I essentially tried to replicate that momentum that IBM's billion dollar investment did to Linux with RISC-V.
Today RISC-V is a movement that can no longer be stopped. The conversation started to change to, to why not use RISC-V as opposed to the early years where it was like, why the hell would I use something other than what I've always been using? RISC-V is the Linux of the processor architecture world. It's the
transformation to an open source platform that allows many, many more innovators and inventors to contribute to the advance of the technology. I think of Arm as a company that has paved the way for RISC-V. It was always destined to act as a placeholder in the history of computing as a precursor to RISC-V. There is an immense opportunity for RISC-V
to be, in five years, the leading ISA for computing. I mean, it seems to be incredible, but doesn't look so impossible.
Now, okay, all right. D, do you want to take over the slides? Yep, yep, I got, I got here, I got one last slide and then we're perfect, perfect. Here we go, oops, yoink, there. Yeah, so, uh, this idea that came up in the video that some of the people involved, I really like the idea of a
lingua franca. Basically, we showed you RISC-V in this course, and basically all the instruction sets are about the same.
It's not like one is gigantically better than the other. So, as was pointed out when we talked to the pioneers, RISC-V if we just got one of them right, we could just all embrace that and we'd have a real lingua franca for instruction sets. And as a modu of, uh, from India pointed out, if it's going to have a lingua franca, it needs to be open. So we think that's our
goal, right? And, you know, by the time you guys, if you're a freshman, by the time you graduate you could see how close we are, or maybe if you go to graduate school, by the time you graduate by then it could be the most popular instruction set. It's certainly capable of going from the smallest to largest things. There's no technical thing involved, no real challenge, but you know there's a lot of business issues to PR happening,
but this would be very exciting if, uh, uh, an instruction set that came out of Berkeley, in part to help teach 61C, became the instruction set that everybody in the world used. Okay, I'm ready for questions, and we have a couple queued up already.
Bor, you want to grab? You know, Dave, I mean half of our labs are now RISC-V and projects in 61C here are RISC-V, but you know, take a look in about a year, a year and a half, I think all four will be, when we have more RISC-V hardware available and affordable, I think we can, uh, we can switch the whole course, and when the hive, I mean, I mean when the hive is all
RISC-V, I, I can't wait till that day, uh, that would be wonderful. Well, there's a project that I'm heading up that's called the RISE Lab. It stands for RISC-V International Open Source Lab. So there's a joint effort between Tsinghua University and Berkeley that's located in Shenzhen that's been around. This
research institute's been there for six or seven years, and we're starting a new project there called the lab, and the first thing we're going to do is build what we call Pico Rio. So this Rios is Spanish for rivers, and so we think of the Rios lab as collecting resources from all over and creating this powerful force like a river, and Pico Rio or
Little River or tiny river is going to be a Raspberry Pi-like board that's uh based on RISC-V, so it'll be trying to be open source as possible, but instead of using the ARM processor it'll do RISC-V, and our goal goal is to make that a lot less than $100, uh so we're going to do that for the first half of 2021 we hope, and so that we hope to supply a a
low something that students could use that has runs RISC-V and do software development. So okay, all right, we got a few questions from the students. Ready? So the first one, this is an interesting one, this was the same question was asked um when Andrew Waterman and were guests over here, right?
answer to this one: what aspects of the currently frozen RISC-V ISA, if any, do you disagree with or would change if you could?
me so I I I mean Krste Asanovic is just you know a real you know he's kind of a mile wide and a mile deep, he knows a lot of things and he knows them really deep and he really likes to micro optimize, so that explains the uh the funny swirled address bits, it actually
it actually does save some hardware answer right right there's a bit and then there's a bit and then it saves some hardware in the very lowest end uh designs, but for me in because of you know the cost is having to explain it, I think I would I would have just done straight address bits uh and which would
have hurt the low-end designs a little bit but it would have been a lot easier to explain I think think that is that is the one that stands out to me. Yeah, I think both Andrew and you are blaming it on Krste. No, I think that's brilliant, I think I think Andrew, I mean it's kind of like I've known Krste since
he was a you know he took classes, I'm old enough that he was a grad student here, I was also Dave's grad student, yeah, and uh but he he just really he kind of enjoys puzzle solving, so for him to kind of play with the address, uh, with the op code bits and, I could align these four instructions with those four instructions and that might save a
couple of gates here and there. So it's like a puzzle solving, like people enjoy.
Do, he likes doing stuff like that. So for him it was kind of fun to play with, muck with the opcodes to get that little optimization. So it was part of his personality. So yeah, Krste did that. No, nobody else would have taken the time to do that that I know. I like doing that too, I have to admit. Okay, well
Borah Borah, so it's like Karnaugh maps or something like that, right? Right. We haven't, we haven't covered Karnaugh maps in this class. We leave it for later. Okay, here is another question. This one is, uh, kind of, you know, we answered that, but maybe Dave can say: why do computer manufacturers not adopt RISC-V now, given
its advantages over other ISAs such as x86? Yeah, so, um, you know, there's a, first of all, hardware is not like software where, you know, you do a startup company and three months later you got a product, right? You have to design it, then you have to manufacture it, and it takes a couple of years. I've been work, you know, I've known industry some, but I've been
at Google for four years, so two or three years to start a project before a product comes out and actually deploy it.
When, when it's a real product, it's got to work. That's typical. So partly it's happening and we can't see it. The other thing is that you guys don't know this but Intel had this problem: how do people know whether this x86s and things? So they did an advertising campaign on television called Intel Inside, just to let you know, right, here's this thing.
It's got, it's called an IBM PC, but actually it's got an Intel part in there and they did advertising for that. So we're not gonna be able to do that with RISC-V. So there's gonna be a lot of stuff that you use that has RISC-V inside. So for example, Nvidia GPUs have RISC-V inside, Western Digital drives
RISC-V inside, Samsung cell phones RISC-V inside. So there's a whole bunch of them that we know have RISC-V in them, but it's going to be hard for for people to notice it. And we talked and and and there's no budget, right? There's not a there's not, I mean, maybe the foundation might decide to do a similar kind of RISC-V inside thing, but it doesn't ask for funds
from companies to spend on advertising, right? No, not happening. Yeah, not not happening.
Yeah, don't wait for it. Exactly. So it takes a couple of years to get out out there, and that's why I wanted to show you that video, what Stefan said. It's gone from why should I take any risk with this instruction set to like, wait a minute, this is working, right? Why shouldn't I use it? So that's going on. The places where it'll be the most
difficult are ones, now that you guys understand this, that require binary compatibility either with ARM or x86. So where does that matter? IBM PCs, the IBM PC binary compatibility. Now on the other hand, Apple, right, abandoned the x86 instruction set like a month ago, right?
Yeah, change instruction set, that's hard for most companies to do. Uh, the other one would be um uh would be the uh uh the smartphones where you're distributing applications in binary in the ARM instruction set. So those two will be tough, but going forward, if you're going to do something
new and exciting that's better than a cell phone, uh, then you know, uh, you can pick what instruction set you want, because kind of binary compatibility is not as important today as it was in the past, and so that lets us innovate in instruction set architecture. And then here comes this new open one that is uh that has all kinds of potential
benefits, and so I think we're going to see, like I said, in the next five to 10 years I think it's going to get very popular and, you know, you'll be able to see, check that. All right, here's another one somewhat related to that. Uh, how does RISC-V plan to supplement uh legacy extremely legacy systems like banking and medical computers? Well, that's a
really hard one. Like the IBM System 360, which dates back to 1964. IBM bet the company on this idea that you could have binary compatibility across the whole family of machines, and they won that bet. Those so-called mainframes, this was before, seven years before the first microprocessor, that instruction set is
still alive, well, and doing a 10 billion dollar a year business. Wow. So, you know, it kind of, you know, it's interesting. Can you capture the existing applications?
But it's more exciting from my perspective to get the new applications, the things that haven't been built yet. If those are going to be in RISC-V, and if, you know, those older instruction sets have legacy code that can last for a long time, that's fine. The other possibility that people talk about is what's called binary translation, where you actually treat the instruction set
as if it's a programming language and compile it into another instruction set.
That's a topic that people have worked on over the years. You know, that might work, but for me it's like, you know, if we're going to do 20 billion Internet of Things devices or 100 billion Internet of Things devices, it'd be more exciting to see these new apps using RISC-V than it is to, you know, kill the IBM mainframe market. Uh, that's alright for me. Right,
all right, here's another question. This is a good one. We haven't thought of that one yet. For the modular extension instruction sets, is there modular hardware to match?
I'm wondering how that general connection works. Well, I think that's, you know one of the exciting things about this transition, the open instruction set architecture allowing open source instruction sets, and then this slowing down of Moore's law, meaning that we you need to start building accelerators to improve performance and performance per
watt, is I would say the module the the the the hardware equivalent of modularity will be these so-called IP blocks. So what I've been doing at Google is building accelerators, I've been helping build accelerators for machine learning. So what we need is like a standard processor that runs like user
interfaces and uh you know compilers and operating system stuff like that, talk over networking, and these accelerators that make machine learning really shine, and that's the accelerators I've been building. So I would say that's the modular hardware equivalent of it. It'll probably be blocks written in SystemVerilog or something like that that you
can buy from somebody else and integrate on your chip, uh, would be the modularity there. There's another idea that's catching on right now with the ending of Moore's law. Moore's law basically, when he made his prediction, by the way, I'm sure you said Gordon Moore, alum Berkeley, right, said that, okay, yeah yeah yeah, that's to a
slide, is uh is that he was betting that every year, or in later revisions every two years, twice as many transistors per chip, that basically was going to happen, more transistors per area, and the chips would get bigger, and they just got bigger and bigger and bigger over the decades. A new idea now that Moore's law is petering out is what's called multi-
chip modules. Instead of having one big chip, it's lots of little chips that are assembled on some kind of substrate. So another version of a modular hardware idea would be this multi-chip version, where you might have, instead of IP blocks, you integrated a one chip like you know typically what we've done in the past, maybe you'll have actually
little chips that you're integrating together in some substrate that's really modular. So you could have a floating point piece, a chip, and you put it in the multi-chip things, or you leave out the floating point chip. So that's that's a possibility of what could happen. But basically it's a very exciting time in computer design, because all the the way things have been designed for decades
doesn't work anymore, and so people are trying a lot of innovative ideas, um, and in instruction sets, the way we build hardware, all kinds of stuff. By the way, this latest idea that U they they've shared about trying to build modular stuff out of chiplets is, um, you know, based on RISC-V. It's Andreas Olofsson's idea.
He just said he has a new startup, uh, he was video, he must have got it for me. All right, here's another question, this is a good one. Are there vector instructions for RISC-V, or are there plans to implement vector instructions like AVX in Intel? Just for a background,
our current project is using vector instructions and intrinsics, uh, uh, to to to do the parallelism. Yeah, so one of the jokes is about RISC-V is that, uh, kind of what caught on is SIMD architectures, and there's a bunch of us at Berkeley who think those are really stupid designs. In fact, uh, Andrew Waterman and
I wrote a a paper called SIMD Instructions Considered Harmful. But the contrast is these vector architectures, which are very elegant. So a kind of one of the jokes is it's called RISC-V, we called it RISC five, it's called V because we wanted to have a vector instruction set. So Krste Asanovic and Andrew Waterman have worked very hard
with a bunch of people in the RISC-V Foundation to finalize a vector instruction set. So it's, you know, it's almost done, people are starting to use it to build it. So yeah, there's a one of the modular extensions is a real vector instruction set, uh, which many of us are excited about, looking forward to. How, how why are they proposing it? It's up to you.
You can, it's, you know, it's built in. That's the elegance of vector architectures, is you get to adjust that.
But I think the people are talking about some of the first implementations are 256 bits per clock cycle or 512 bits per clock cycle. I think people are in that range right now, but you know that has long legs, it could be whatever people want to build. That's exciting.
Yeah, it's, it's, you know, it'll be another year or two until we get hardware, and then, you know, Hive will be replaced and um, um, here's another good one, lots of discussion from projects here: how similar is the way people design CPUs to our Logisim project? And they remember
Logisim, it got better. Yeah, I remember. How similar?
Um, well, people do design it in a hardware description language, they, they absolutely write it out. That's Logisim is a, is a hardware description language, right? Yeah, I mean it is a graphical interpretation of, of, uh, I maybe I can answer this one, why don't you answer it.
Actually, this year we built a new back end for Logisim. Logisim now exports Verilog, which is a standard language that people use for building CPUs. And you guys don't know, but we, when we do, when did these project competitions, we actually went and synthesized your CPUs. We didn't tell you that we did that, but you actually, there
does exist a live, you know, you submitted a Logisim file, but in the back end we actually had the hardware description, real hardware description of your CPU, and so we were able to compare them.
Uh, this is, so the, the embarrassing thing is Verilog is a kind of a terrible hardware description language, but it is popular, uh, it's inelegant, so that's, um, there's a lot of reasons not to write in Verilog, uh, there's a variety of, you know, even in industry higher levels to generate it. Verilog is kind of like
almost writing assembly language, uh, way of designing it, and so a lot of tools use it, but if you wanted to get an idea of the low, you kind of the worst case scenario of what hardware design descriptions look like, you look at the output of the Logisim to see it, but I think if you've done Logisim and then you start, you could, uh, what if you take a follow-on
course, I'm sure you could pick up various hardware description languages and understand it. Yeah, these questions keep flowing, um, at some point. Great. No, no, we're doing, we're doing fine, we have, we're fine. Okay, uh, and they're good, I mean, you know, they're they're demonstrating. Yeah, right, that's
right, crickets, crickets on their chat, on their chat. Yeah, they keep going. Um, here is a, um, here's a, you know, I'll leave the last one, a second part of this one, for for the end. Um, first from Riata: while technical question, would you or the RISC-V Foundation consider making
instructions for quantum computing as it is getting developed? So, uh, because I give this talk about, you know, uh, things are slowing down and we need innovation, I get asked a lot about quantum computing. So quantum computing is a futuristic thing, there's a couple of national reports that talk about that
maybe it'll, uh, have applications that'll work in 2030, and keep in mind that quantum computing is big computing, it's big science, right, near zero degrees Kelvin, trucks going by make it not work, so it will be like in a data center if it does work. So it's a really interesting topic. You know, when I
started graduate school, there were a bunch of people interested in nuclear fusion, they thought, I'm going to spend my career on nuclear fusion.
Nuclear fusion still doesn't work, if it ever works it's going to be really fantastic. Uh, we know it works, we look up at the sun. Yep, that's nuclear fusion, but we haven't figured out how to put less energy in to get more energy out than we put in for nuclear fusion. So Quantum confusion is very exciting. I don't think it's going to be practical this decade, so it's a 2030 thing, but you
know, if you're excited about it, you know that's really great. So but I don't think it, it kind of doesn't make sense to have a Quantum Computing instruction. That's not the way Quantum Computing works, but there are people here you could talk to if you're excited about Quantum Computing. I'm sure there's many projects, uh, you can learn more about it. Yeah, here's another one and a
few related ones. I mean this is, you know, there is a train of, uh, you know, you mentioned this, so uh here are the questions related to that. Apple's new M1 chip is Arm based. Do you see the possibility of something analogous based on RISC-V and how, why see something what, say again, analogous based on RISC-V?
No, I would say, um, this happened some in the RISC-V Community where somebody had a processor design that I said kind of these RISC architectures are all pretty similar. They had an existing design and they learned about RISC-V and they liked the idea of this open this and this community around it, and so they just changed the, you know, the first stage of the pipeline, the
instruction, well after you fetch the instruction you decode it, the second stage of the pipeline, that decoding. They changed the decoder and they could execute RISC-V. So if they can, that M1 design that's based on the Arm instruction set, they could pretty easily switch it over to RISC-V. So I think whenever Arm is successful in building something that's faster and lower power,
to me that's a great sign because I'm sure you could also do it in RISC-V.
five. There's not, there's not, like I said, the instruction sets are more or all more or less the same. x86 is very different, but the RISC architectures: loads and stores, you know, 32 registers, you, uh, similar encodings. So they're kind of all the same, and you know, that's partly, I think, for Hennessy and I, we got our Turing Award in part
for our contributions to RISC and in part for that textbook. Um, so we're kind of surprised that 35 years later the best ideas are still RISC. You know, 35 years of Moore's Law, you would, who knows what would have happened, right? We're kind of surprised, but since that seems to be our best ideas there, uh, any time something good happens to ARM we think
that's good for RISC-V, and assuming that people, some, a significant number of people will like the open model as opposed to the proprietary model.
And so here is another, you know, this is another good question, and you know, there is another person who I'm thinking of who could be, uh, who could answer this, one of your former students. Can binary translation be done in hardware? Yeah, so yes, as I think you know from the class, Intel, to its credit, once the RISC ideas went out, didn't give
up. They put a 500 person engineering team together to build the Pentium Pro, as it was called, that came out in, I think, 97, and what they did, amazing enough, is translate from x86 into RISC instructions on the fly in hardware, so kind of a hardware compiler. Now at that time Intel's process technology was a
couple of years ahead of everybody else, so they had extra stuff on the chip, but their technology was years ahead, so they could, they could hide that cost. So there's an example of somebody doing hardware translation on the fly. Now there may be some, in one of these discussions, uh, talked about it. There somebody claimed there were some legal
issues about translating, uh, binary instructions and executing them, and so I don't, that like patents, whe- whether you'd be violating some patent if you did that. I don't actually understand that, but I have heard that claim. So before you start your company, you should talk to some lawyer and see if there's some truth to that uh rumor that there may be, um, you
know, you may be violating some kind of patent if you were to do a hardware translation. Yep, um, I I actually know of a patent, I think, that that claims that um yeah I we are getting so uh two two questions here. Um, one, you know, I'm going to rephrase some of these because we
have still quite a few and we are running long on time, and I'm going to save Reiot as a second part of our question for last. The one, so what are the plans for these, for your books? Uh, I mean I I actually, you know, my age, well I'll reveal my age when I tell that I actually own a first edition of the book. That's how I learn. You're you're
really old. You're really old if you if you bought a book. Yeah, I actually own the first, the second edition, then like I skipped the third maybe. Yeah, that it wasn't very good.
Okay, what's my, what's your question? How many total copies do do you keep track of how many total? We asked once and I think uh the undergraduate's more popular. I think the graduate book was, I think it was a quarter million to a half a million, something like that, and the undergraduate one is, they certainly sell more, there's more undergraduates than there are grad students. Over this is
over, you know, 30 years, right? So it's a number like that. So yeah, hundreds of thousands. Yeah, wow. Um, and um are there plans for a new oh absolutely. Uh this we've we uh the undergraduate book, uh I just turned in
the G for the second edition for the RISC-V version. I talked to Krste Asanović about it, asked him based on teaching 61C. What, what did he recommend? He recommended going back to 32-bit addressing because, you know, it's, it's when you're first learning your instruction set, there's that confusion about 64 and 32, and there's a few
instructions there. So we did that. Uh, I also don't tell them about the swirled address bits until chapter four. I just, I don't explain that.
In chapter four we explain what multiplexers are, and I have a little table there that shows exactly how many three-to-one muxes became two-to-one muxes because of the SR address bits, and I there may, there may be a sarcastic tone to that section.
But, uh, and we emphasized more about, uh, about fundamentally the bits, bits can mean anything, you know, that, that it's kind of a strange concept that data can be anything. So I try and do more, kind of hit that one harder because people are surprised about that. So these are a lot of suggestions from, uh, from Krste
for the second edition. So I just turned in the galley, so it'll be for sale maybe in, in March or the next edition. All right, next fall definitely.
Yeah, and here is one final question: uh, what are a few words of advice you would give to folks who are just beginning a college CS career, and maybe we can, uh, expand that or narrow it down a bit to people who are just wrapping up 61 C. Yes, uh, so I, I, as Dan and Bora know, I
love giving advice, so you, you're gonna have to cut me off probably. I would say, you know, I'm sorry that you're going to college in this, in this quarantine time. Uh, I've been around a long time; this has been the strangest year in my lifetime. Uh, so, uh, you guys, uh, survive that. I, I, I think you picked a great
major. If you're, you're in this to do computer science or EECS, you know, I, this is going to be a really exciting century, just like the last century was really exciting. When I, there was no undergraduate degree in computer science.
When I went to school, uh, that happened a couple years later and I kind of did it myself, I stumbled on this. But, uh, you know, my uh, when I did it, my relatives said, well, okay, good luck in computing.
And then over the decades, wow, that was really smart decision you made. I did it because I loved it, not because I thought it had a tremendous future, but it, it worked out really well and I don't see anything slowing down in this decade. Um, you know, machines starting to get more intelligence, more applications, it just seems like this is going to be a very exciting field. I, the part of Google I
work at is called Google Brain and that's one of the big proponents of the machine learning, the deep networking. I would definitely take statistics courses while you're there. That's, you know, computer science and statistics is the foundation of this, uh, machine learning.
It's hard for us to know how, how big a revolution this is, how far it's going to go. It's already gone pretty far. Uh, it could change how we program. Instead of writing millions of lines of code, already there's some examples where we in, er, those lines of code basically from algorithms that look at the data. So I
think it's going to be a pretty powerful force. You should think about going to grad school, right? Uh, life expectancies are going up. Uh, I, I retired and one of the best things about retiring was going to a retirement party and she said the number one correlation with longevity is education, and so people in this crowd are going to
live a long time. And I had been thinking, based looking back, that boy, I'd be lucky if I get into my 80s. And talking to, I hear, I added a decade to my expectations. I expect to get to my 90s, mid 90s. You guys, a lot of you will live to a hundred, being get to a 100 with advances in medicine and taking care of yourself. So that's a long time.
So thinking when you're 21, boy, that's it. I can't afford to go to school anymore.
You're gonna, you're gonna have really long careers. I would seriously consider graduate school, the last degree you can buy. No one's got a PhD plus, is there? There's not like PhD plus, right? And spending five or six years getting acquainted to all kinds of the technologies that are going on, people still willing to teach that to
you, and you can. It's a ripe environment to learn, and you get trained how to do research on your own, give talks on your own. Part of the reason I ended up at Google is I wanted to do a sabbatical there, because I didn't understand why they're hiring all the PhDs. And more or less they said, look at the training you get with a PhD, you can tackle problems by yourself, you know how to solve them, you learn how to communicate. Those are
valuable skills, so they hire a lot of PhDs even though they don't have a big research branch. I think going forward, given life expectancies, uh, PhDs more, there's going to be more opportunities for people with a PhD, so you should seriously think about grad school, at least a master's degree.
You'll have plenty of time to make money after grad school. Don't worry about that. All right, great advice, wonderful.
All right, thank you, Dave. Thank you so much, Dave. It's great to see Patterson, everybody. Hey, I can see you still doing, you're still doing competing in the news.
I think I started that. You did start. Thank you for that. Right, wonderful.
All right, take care, Dave. Beautiful. All right, let me share my screen and get it here. That was so, what a delight, Bora, what a delight, my goodness. All right, so I think it's this window, see if that's going to work.
Okay, let me try it again. Okay, here we go.
Yoink, we got some news I'm to share, this is the midterm scores, so excited about that one. So first, start, first we'll start with the computer in the new. Yeah, yeah, exactly, this is a good one. Yeah, 75, so Here we go: Tesla Model X gets hacked through a new relay attack. Tesla said it's pushing a patch through software updates.
By the way, this is one of the benefits of having a software, you know, an over the air software update. Uh, normally you have, if you have a, there's a software failure, you have to, like, queue up a time.
Well, I'm Subaru, or BMW, got to queue up a time to get back to the, to the dealership. This is wonderful, you can update this there, but I, I also understand, I was reading the article, um, that talks about there's a two-step hack.
One of this is the, there's a key fob. Did you read it, chance to read this? I was reading this while you were talking to Dave. Um, there's a key fob that's used in the, in the Tesla, and they were able to basically spoof the key fob to open the door, and then the catch is you can't just drive away. You can certainly, with the key fob, I think, roll the windows down, open the trunk and open the door, but you can't drive away. They say, however, what
they do is, inside the vehicle there's like a little port. You know, many cars have the little port where you can control, you know, plug a laptop in with a serial port and talk to it. Within, like, a couple of minutes he plugs into the port and he overrides, he says, yep, okay, you can start driving, drives away. And there's a whole video on this, on this link. So, uh, certainly Tesla cares about security a great deal, and they're going
to be trying to push out a fix very, very soon to fix this. This does show a few really important things that, uh, again, somehow we got this back into RISC-V course, right? Uh, you know, the, the openness and understanding where, you know, building in security in the very root of our systems is going to be really, really important going forward, because, you know,
yeah, you know, there is a great comment over here: I liked it when the most advanced thing in my car was the radio. Cars, cars, I used to own, oh, I didn't even have a radio car, but the point here is yeah, this is
uh, the big thing here is we like the convenience and we will love to talk to our cars, you know, coming, you know, in a few years we'll be talking a lot to a lot more to our cars than we than we are presently, and there will be a computer and that computer will need, or that set of computers will need, to be very secure.
Um, that is, there is a big debate over here how open this thing has to be, and RISC-V is making a really big push, right, that uh um that uh things should be completely opened down to the last single bit, and then it's a lot harder um to to to you
know, if there is an attack, then it's very easy to trace it down and understand where it came from. And I mean the topic about here is cars, obviously.
We're just trusting that Tesla does the right thing, right? We putting our lives on the line to make sure Tesla does the right thing in terms of security, and so car doesn't get stolen or hacked or you know somebody puts the brake on or accelerates it during a traffic jam. On the other side, we're talking about uh voting machines, right? We just had a big election a couple of weeks ago, and so many voting machines are still closed systems. I think the future of voting machines, and David Wagner, our own
colleague, is working on that, uh, is thinking about could you have a requirement from the federal point of view that says all voting machines need to be open, open in software, open in hardware, so that anybody can look.
Obviously paper trail, we talked about that, always critical. You Bo you get paper trail, then you hand the paper in as your vote, and maybe the voting machine is just, you know, you user interface to make voting easier for somebody.
But I love the idea of of mandating from the federal point of view an open hardware and software system to all voting machines. I think that'd be, I mean, it would certainly put a wrinkle in many companies who already have investments there, but I think that's the right way to go moving forward. It's an opportunity for the next set of companies. Amen, exactly, exactly. All right, so I wanted to share a midterm score
update. Uh, Bora and I have written a midterm, and the midterm we wrote, we wanted to do better. Uh, we were really hoping for a better number than this histogram, but this is what we got.
Uh, as you see, it's a very bell-shaped curve. Uh, you know, unfortunately we've got a median of about 50% and mean about 50%. And this is, again, this is Bora and my midterm, this is what we tried, and we tried to write this, and wait, Bora, did you, did you play the slides here?
Something looks wrong with the slide here, you know. Yeah, this, the age maybe.
Yeah, this is weird because, because yeah, the year, this is fall of 2018, the last time Bora and I gave a midterm.
Oh, this was the midterm. Okay, are you ready for this? This is what you get, ladies and gentlemen, the new school: Bora, Dan, CeCe, Justin, Jerry, Stephen, everybody.
W!
83%. Clobber, baby, clobber all the way to the bank. I love it. Was that good? Did you like it? Did I? That was, they're traumatized. Is that okay? Did you guys crush it or did you crush it? Oh my goodness. Okay, this was not a lot easier, fair, this was a fair, a lot easier
midterm than what we had before. This was a very fair midterm. I think that it wasn't like we just gave it to you, you know. This is, you guys did a great job, you helped each other, you guys are amazing. I want to tell you all, please do share this with your, you know, link to this part of the video when you share this with your friends. Make sure you don't just read the answer. We're going to post this in about, actually just posted. Now we just post the grades now,
and we're not doing the silly joke about the old school, the new school. We just post the normal grades. It's posting right now, it releases right now. You guys crushed it, crushed it, crushed it, crushed it. We are so proud of you all. Great job, everybody, studying so hard, working together through the, you know, but through the COVID, through the pandemic.
You guys crush 61C. You think you're awesome. Great job to everybody, really, really wonderful. This is really a testimony.
You know, we streamlined a few things and they are going to stay forever with this class, and this just shows that I think that the material went through. I mean, yeah, and we were seeing this on the project scores bar. I mean, this is not inconsistent, right? The consistency that we're saying is you guys are doing an amazing job this semester despite everything, right? Despite all the COVID and
all the this and all the everything we're doing. We're trying to make this a better class, and I think clearly from your midterm, you guys are doing amazing.
It's supposed to release now. I'm seeing some people saying they didn't get it. Uh, I'm really proud of everybody here, proud of our team, I'm proud of our staff, I'm proud of everybody involved on the staff part of view, and proud of the students. Now that said, some students are still struggling, right? We can't always celebrate unless you all got a 45. We're not, we're not done yet, okay? We still have work to do. Please come talk to our
staff if you didn't do so well, people on the lower end. Oh, by the way, this is very important, we didn't tell you this, this, this number includes all the people that had to do a retake for whatever parts they have. We put them, we had to put them the way this, the grades were. We didn't, the software didn't know that they did have to do a retake, so they have zeros. So a lot of people have zeros or low scores, they're going to be
totally replaced with the retake score.
And that is still in process, we're hopefully getting done in the next couple days, so you guys, even if you got a low number, don't freak, because it'll say, by the way, I think we were able to put in in the Gradescope, it'll say to be replaced by the midterm retake. So don't panic if you got a zero, because you knew, because the retake isn't done yet. This is before the retake, so these numbers are going to go up. So 83 is the lower bound, we're going to be crushing
this, this histogram once we get the retake scores embedded into here. So this is the beginning, this is the beginning of a conversation before the retake comes in. It's going to be amazing, so great job everybody, uh, in terms of this, really, really well done. Let's do a quickie schedule. I know we're, we're running late, uh, we're at week 14. There's no lecture on Wednesday. Happy Thanksgiving, everybody. No lecture on
Friday, uh, spend some time with your family if you can. If not, if it's on Zoom or if it's remote, please do spend time and be thankful for all the things we have, uh, in our society. We don't have office hours this week, correct? No office hours this week, all the Wednesday, we, we basically shut down after Tuesday, we shut down after Tuesday and we revisit, and we try to make it so that you can really take the time to, you know, enjoy
your time. That's why we moved P4 to 12/2.
So nothing really has to happen, hopefully, hope you started on Project 4 already and you're making progress, but you can really take some time to recharge and reg- get some energy. Um, next week, Stephen, want to remind me this? Um, next week we've got the normal, uh, you know, recorded lecture. Borivoje did a wonderful job on dependability, parity, ECC, and RAID. That's, it's already posted, beautiful, already posted. That's
that's his last lecture. I did my last lecture on cloud computing, which is today's lecture. Then this is important, we're gonna have a wonderful guest from Apple, uh, I want to get the name until it actually gets finalized, and and we're we're double, double, triple check that, but we're hopefully to have a guest from Apple talk about a GPU in that lecture, and then he's going to be giving another lecture on the 29th live, so we think, or
we think it's going to be live, so we're going to probably end up, um, using this exact time slot, which is the five to four to five time slot for a live lecture on GPU. We're very excited about that. We're g hopefully announce it formally on Monday, and then please do come live on 12/4. We're going to have the last session live. We're gonna again cannibalize this webinar to have the
last section live. We'll have all the awards, uh, for all the prizes, all that's going to be done live, so please do. Next week is recorded lecture, Monday live, we think it's going to be live on Wednesday for GPU, it might be pre-recorded. If so we'll let you know, and live on, definitely live on Friday 12/4, so please do come for that, uh, and please note that
the Project 4 deadline is 12/2. Don't forget that. In in the in the last lecture we're going to have Project 3 prizes, we already have them. Uh, Project 4, we have to wait for all, I mean there are people who have various kinds of extensions. Yeah, we have at least six slip days, we have to wait till so Project 4 we're not going to be able to announce by exactly few extra probably. Yeah, well, and plus we'll need
time, like once your once your slip days are done, then we have to run our project competition, that's on our side, so we need to be able to make sure that's checking out. Okay, by the way, if you're missing us, I mean we are we are still [Laughter] around. Yeah, and Bor, do you have any room for undergraduates in your research group? Because I S we're going to advertise that last week. Yeah, we'll we'll advertise, asked my colleagues, we
have a ton. I mean, worries about that.
Good job. All right, and I'm hearing that scores are just released, so you guys are all good. I'm so proud of you. Hope that's a great gift for you to have a nice Thanksgiving. We will see you next week on Monday for our live session, looking forward to that. And don't forget, come to we think it's going to be Wednesday live session and definitely Friday, so circle that in your calendars. So we'll see you there. All right, guys, this is a rare
thing, Apple speaker, you know, public. Yeah, how to design the GPU. We're excited about it, we're excited about it. Good stuff. All right, thank you. All right, thank you all. Have a great time, enjoy time with your families on video or live and take some rest and come on back rested and recharged for our final exam.
We'll have our uh final exam uh details released uh on Monday 11:30. Okay, thank you. Take care, folks. Happy Thanksgiving, everybody. Good to see you all. All right, take care. All right, all right.
Đăng nhập để đánh giá học liệu này.
Đăng nhập để bình luận — chưa ai bình luận về học liệu này.
Không có khóa học liên quan.