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.
BORIVOJE NIKOLIC
Great question. Can I um — PROFESSOR: Yeah, perfect.
Wonderful.
Okay, and students are dropping in now. Welcome, everybody, to week 10, 61c.
BORIVOJE NIKOLIC
Well, it looks like at least some people have power.
PROFESSOR
Yeah, that's right. Yeah, that's true.
BORIVOJE NIKOLIC
I don't. I'm in my Cory office. This picture is fake behind me. It is fake, but it's almost real.
I made it to campus because there is no power at my home. We have a limited amount of time.
We have something really, really, really special today, so we're going to get going to that.
We have a CTO, a well-known person in the industry, which is Mark Papermaster from AMD.
And he will tell us some history, how to solve big problems and have fun in the next 20-ish minutes. And then, feel free to type some questions.
And we have a two — a corporate fellow Mike Clark, and a senior fellow, Tom Burd, who can help field some questions afterwards.
Mark, take it over.
MARK PAPERMASTER: Very good. Dr. Nikolic, thanks for having me join your class today, and it will be quick. I'm going to go through a little material, but it really is about engineering.
I'm am old guy here. I've been 35 plus years in the industry.
So I'm just going to jump in and just give you some thoughts, and share with you just some of the experiences I've had. So it'll be a little mix of things here.
Bear with me as I just get my bearings here. I'm addressing the slides here in Zoom. But look, time goes by pretty quickly when you're having fun.
So the first thing I'll tell you is, as you're in your studies and you're deciding what do you want to do in the industry, what do you want to do in academia, the biggest thing I'd say is it's fun to solve problems.
When you're solving problems, 35 plus years can go by pretty quickly, because it really is fun. And I've had a chance, over those years — I've been primarily at IBM.
I'll give you an example of IBM, and one of Apple, and one of the last nine years I've been CTO and running technology and engineering at AMD.
So a few examples of how you can both work really hard and have a lot of fun. So I'll start. Look at this photo that I'm showing there.
It's an IBM supercomputer that was part of what's called the Accelerated Strategic Computing Initiative, where rather than blowing up atomic bombs, the intent was to stop that practice that was going on across the world and simulate them.
And so we created a supercomputer. And it's pretty amazing that this was 10 teraFLOPS of computing.
So it took a lot of chips and a lot of systems to create this type of computing back then.
And so it was a lot of fun, though, because we really had to boost the performance, and really think about problem solving at scale, how to scale out.
So it really, really brought the team together very, very closely and we solved some tough problems. In fact, this particular one called ASCII White was the first use of copper interconnects.
So it was not only the designing of the computer chip, but how to use new materials. And that's quite common in the chip design industry.
You'd think, well, I can just work on one aspect of it, but not true. It's really getting many disciplines together.
And that, after some years at IBM, I took the job running iPod and iPhone at Apple.
And that also showed me the incredible power of collaboration. Because, in fact, you look at the new iPhone that just launched — a lot of similarities to this iPhone 4 when it came out.
It had retina display. It had glass front and back. It had CNC around the metal band on the entire periphery. It was the first introduction of FaceTime.
It was a new high-resolution camera. So new, new, new, new, and a new hardware and software stack. And so it really took a phenomenal collaboration on the goal.
Everyone focused on, what's the end goal, and working all the way through not just the internal engineering teams, but just like that last example I gave you, all the way through the supply chain in the manufacturing side.
And so it's one of the things I highlight is, solving tough problems really means coming together and anticipating really what the use case will be, and how you have an end product that really is great and changes experience.
So iPhone 4 was certainly a great example there and a really exciting program. And it brings me to Zen.
So you have Tom and Mike on this call, who are Chief Architects. Mike is our Chief CPU Architect at the company. And they can tell you a lot about what we've done with Zen.
And we couldn't be more proud with launching our third generation of Zen, which we just announced.
And then we begin shipping it in just another week and a half, so very, very exciting time for us. But that was another huge challenge.
You've got to call a problem very candidly, and our problem was we had let our single-thread performance slip.
So the team rallied together and set the goal of getting AMD back to performance leadership.
And it took really setting that right goal, really defining, what are the knobs to get there, and delivering that.
And so we did that in 2017 with the first generation of Zen. And I'll talk in a minute in a little bit more detail about the Zen 3 that just went out.
But again, each one of these stories — it's not where one person just went off in a corner and could just innovate and come out with something that created a great product.
It really is about teams, working together, solving problems together, and making a big difference. And from our standpoint at AMD, we've got to stay on that.
Because you look at the trend of GPU and CPU performance, it's been doubling about every two plus years, and that must continue.
The demand for more computing is insatiable, and we don't get the lift on every semiconductor process node that we used to. The transistors really aren't getting faster.
Some of the circuits aren't scaling as you go node to node. And frankly, they're more expensive per transistor. So it's just a phenomenal time for innovation.
Because any time someone says, oh, because of this factor, that factor the performance gains are going to stop, they're just wrong. Any time you hear that, you can just say, look at the history.
There's always been innovation that keeps us driving forward and providing more and more performance.
And certainly, and as far as I can see, there's just no end in sight as to how we're going to stay on this curve.
So I'll just take a minute and just talk about graphics, rendering graphics.
When you're drawing a triangle it turns out there's a phenomenal amount of computing, because you're rendering billions and billions of these triangles to make incredibly resolute images.
I mean, the whole idea of any of us that are involved in graphics development is we want it life like. We don't want you to know it was even a rendered image.
And I'll just show you the example of connecting your studies to having that impact.
In your curriculum you probably had to take the linear algebra, and a lot of the math — math is the language of what we do as engineers.
And you couldn't see that more across all of what we do to give you that beautiful visualization that we focus on in our discrete graphics products, whether anything from a rasterization, to the geometries redrawing, rendering, and right up to and including ray trace, which I'll show you an example in just a minute of our latest in ray tracing.
Incredibly lifelike images when you apply the math and you combine it with incredibly high-performance computing. So with that, I said I'd show you a little example.
Let me see if I can get this to roll here.
[VIDEO PLAYBACK] [MUSIC PLAYING]
[END PLAYBACK] So I'll stop it there, but we worked very closely with the teams on Terminator in the newest edition of the Terminator series.
And what was out there just couldn't get them the kind of images they want.
And I just show you this in the trailer because, of course, it shows you how you truly break down these images.
But it's a game changer with the kind of CPU and GPU performance that we're able to deliver now. And what you're seeing on movies like this is the whole set's changed.
It used to be that you would film on traditional media and then you would cut, and you would then do another take. And this editing process, this post-processing was incredibly laborious.
And frankly, the directors didn't always know they could even get the right scene. That has just totally changed. This is all done in real time. The processing is done real time.
The editing can be done real time.
So the kind of performance that we've been able to bring is truly, truly changing how the media and entertainment sector is going about delivering their products.
Really exciting.
So just another example of that, and that last example didn't have ray tracing. I didn't have time to show you my ray trace example. But guess what?
Because we have our technology in the new game consoles coming out with Microsoft Xbox Series X and S and PlayStation 5, pre-orders are out there, and I know our family is a big gaming family.
So I've got a couple of members of my family very excited about this. Because it does have incredibly realistic imaging, including hardware ray trace, which just — it's amazing.
Think about it, the games that you played, my god, I think about just eight years ago. The players would be wearing skull caps. Why?
Because the physics of calculating the flow of the hair as the wind blew, we couldn't even render it without it looking very choppy. And so the game producers had, again, have those skull caps.
Now, of course, we modeled those physics. You see the hair. You see flowing robe. You see whatever I'm at, that imagery that they're trying to share.
And now, with hardware ray trace, it's just incredibly realistic, as the shadows and the light, in fact, truly mimic the math of real light bouncing around in our day-to-day lives, very exciting.
I'll now shift over to CPU computation. This has been just near and dear to me. I started way back on RISC processors so many years ago.
And it's just been phenomenal to watch their development, and been such a pleasure to work with a brilliant AMD team the last — over nine years.
And so one of the things that I always highlight to folks is, you don't get lucky in this business of high performance. You design your way to high performance.
You have to really spend a lot of time on the high level design.
If you look at this Zen 3 that we just announced and we just began shipping shortly here, I would always love to just stop by the engineers' office when they were back at the early ideation phase — we call it pre-high-level design.
And on the whiteboard would be idea after idea that could drive performance. And, of course, about half of them don't work. But that's what good design is about.
It's about trying things and accepting failure, accepting that by no means will all of your approaches and all of your ideas work, and throwing out the ones that don't work and moving on, and really testing the feasibility.
And then, when you really have analyzed that design and know you have a winner, then move it in to the phase where you implement.
You do the logic circuit design, the verification, and the validation. When you're done it just has to work.
The mass set with what we're at and these advanced technology nodes is over $15 million per mass set. So that decision to say, hey, I think I got it.
Let's go build this. That's a $15 million decision, on top of all the years of work that led up to completing that design.
So really can't emphasize enough the criticality of the design, and what it is you're studying, and how to be rigorous in the application to come up with great designs and leadership designs.
So that is what we did with Zen 3. If there's time later, Mike and Tom can answer more detailed questions you have. But just to give you the sense, what did we do?
We're coming off of two very impressive designs that drove with Zen 1 in the 52% improvement of instruction per clock, then another 15% improvement with Zen 2.
And so to drive yet another double digit performance improvement, it really was going across a broad range of levers, really how you can bring memory in closer to your compute elements critical.
So improving the latency was key, particularly on gaming.
Games tend to have a dominant thread, and that dominant thread, when you can lower the latency, just, obviously, it's spending less time waiting for the data. And you just see it.
You experience that. So when we doubled actually the accessible L3 cache, and so we reorganized how the cores were laid out.
That, along with really a number of other elements across the load store unit, improving that, improving our prefetch. We always think about it in terms of a balanced computer.
So as you widen out an issue rate, as we did across our floating point, and our integer engines, we had to make sure that not only did we bring a data enclosure with lowering our latency, but we wanted more efficient branch prediction.
And so really pleased with, really, a zero wait time, in we call zero bubble, is really improving how we predict the next instruction and keep a flow of data processing through our execution units.
So we can't let up. It's a very competitive industry we're in. So we're going to stay on this very strong pace of performance improvement. And it's got to be energy efficient.
Tom Burd would be disappointed if I didn't mention as well how we focus equally on the physical implementation because we're power limited in everything we do.
And so what we were able to do with Zen 3 is pretty amazing.
Because at 19% instruction per clock was in the same technology node we had been in with Zen 2 and we did not raise the power per socket that we had.
So really quite an amazing, amazing feat by the team. And then I'll wrap it up with my last section, because I know we just have a few minutes here, but supercomputing.
I've been involved, as you saw from that slide I showed at the beginning of the big Department of Energy computer.
And it is incredible how we have driven what is, like my example before, the teraFLOPS that used to fill a room into a single chip.
So this is our Radeon 5700, our Navi design with 5.4 teraFLOPS of peak double precision.
So GPU is critical for that rasterization and that content creation that I showed you earlier, and equally critical to be paired with CPU and drive, incredibly dense floating point computation.
GPU is the most flexible and easier to program, dense floating point and vector engine in the industry, and we need it.
We need it because you have a traditional supercomputer that needs high-performance computing, molecular dynamics, and fluid flow analysis.
And think about the exploration of the Earth geography for resources — those are traditional HBC that you see on the right.
And you can see that we're going to be breaking 1 exaFLOP next year. In fact, I'll talk in a moment of — we're very proud of our role on that.
But what you're seeing also is a convergence of traditional supercomputing for HBC with the needs of machine-intelligence training.
It's pretty amazing when you look at what's going on with natural language generation.
So we'd all think, ResNet 50, and image recognition, and getting a correct interpretation of those images.
But when you look at what is going on with OpenAI and NLG, it's just phenomenal what's going on with natural language generation.
And to get the kind of accuracy you see there, what they're really finding is you actually need supercomputer class computing. So, again, I mentioned that we're on a pace to an exascale.
And our role with that is that we are the CPU and the GPU underneath. What would be the two top supercomputers coming out with the US Department of Energy.
So it's Oak Ridge National Lab, a 1 point greater to 1.5 exaFLOP machine, which will be powered up next year.
And then beyond that, in 2023, our greater than 2 exaFLOP machine in Lawrence Livermore National Labs, not too far from where you're at.
So we are very proud with both our CPU and GPU to be partnering with Cray, with Hewlett Packard Enterprise, who has acquired Cray, in creating these incredibly efficient and massive-scale supercomputers.
Also at Lawrence Livermore Labs, we have already running today the Corona system. It was not named the Corona associated with the pandemic. It was ironically already named Corona.
But we are powering these supercomputers, and they are indeed being applied to pandemic research.
And in fact, we've just donated an additional 12 petaFLOPS to help further build up this machine, as well as donations we're making across other academic and research institutions across the world.
It is a phenomenally challenging time right now, and all of us have to do what we can in industry.
We're thinking of any way we can bring computing to bear and to speed these type of treatment approaches which are being explored.
Okay, well, again, I didn't have much time so I need to wrap. Look, keep a culture of innovation. You play your role. Use things like Zoom.
I know it's hard when you're not fully on campus. But I can't say enough about leaning into your studies. And really just to wrap, saying it's so exciting what's going on now.
There is no barriers. It is all about innovation. And hardware, I say it's cooler than ever, because it is even more about system design, bringing novel hardware and software techniques together.
It's going to be heterogeneous. It'll be CPUs, GPUs, specialized accelerators. So there is so much challenge ahead of you.
And just always, just be sure you're having fun, and that you're learning, and you're applying new concepts.
And with that, you're going to have just a tremendous career ahead of you in whatever you choose. So thank you so much for letting me spend a few minutes with you here today.
And I unfortunately have to run, but we've got some of the team that can stay with you. Thanks so much.
PROFESSOR
Thank you so much, Mark. Outstanding. Let's bring back Bori into this. Let me do that real fast.
Perfect.
BORIVOJE NIKOLIC
I'm in somewhat.
PROFESSOR
There are a couple of questions. Do you have a second for questions?
BORIVOJE NIKOLIC
I think the others are going to field those.
PROFESSOR
The others will answer that. No, no worries. No worries. Very good, thank you so much.
BORIVOJE NIKOLIC
Mark had a 4:30 meeting.
PROFESSOR
I'm sure he's a busy guy.
BORIVOJE NIKOLIC
He has to be. I think you guys have a financial report tomorrow or something like that.
TOM BURD
Yeah. MICHAEL T. CLARK: We do, we do.
BORIVOJE NIKOLIC
That's kind of important these days.
PROFESSOR
Wonderful.
BORIVOJE NIKOLIC
There are a few questions over here. One was directly to Mark, but — MICHAEL T. CLARK: Yeah, we can't answer that one.
PROFESSOR
You guys can see these too, right? Yeah, you guys can see these, too.
BORIVOJE NIKOLIC
I tested one. I think this is very close to — I mean, maybe we can go ahead and introduce Mike and Tom first. MICHAEL T. CLARK: Yeah.
Do you want me to go first, Tom?
TOM BURD
Yeah. This is on Zen 3? Yeah, why don't you go first? MICHAEL T. CLARK: No, I was going to say, I was going to introduce myself.
TOM BURD
Oh. Yeah, why don't you go ahead? MICHAEL T. CLARK: Yeah, I'm Mike Clark.
I graduated from the University of Illinois in 1993, went to work for AMD on the original K5, if anybody has seen that one. It was AMD's first.
BORIVOJE NIKOLIC
[INAUDIBLE] MICHAEL T. CLARK: It was AMD's first attempt. We were just a second source for Intel chips to the 486. But the K5 was our first grounds-up design.
Learned a lot from that, obviously, our first attempt. I like to say you learn a lot more from your failures than your successes.
And a lot of people may view K5 as a failure, but it was a huge success for the company to push us to get us going and get us to where we are today.
Since then I've worked on pretty much every CPU, K7 K8, if you've heard of these.
Did work on the Bulldozer line, and along the way I actually got my master's here at the University of Texas, very painfully, one class a semester.
Yes, I was still working. But yeah, and then I got that in 2003.
I was a Lead Architect on Zen and now I actually drive the whole roadmap going forward. So I'll let Tom introduce himself.
TOM BURD
All right. Thanks, Mike. So my name is Tom Burd. I graduated from UC-Berkeley.
I did undergrad, master's, and PhD there. So I was there, boy, for over three decades — started '88 and left in 2001.
After, tried my hand at doing some independent consulting for a few years. I've been at AMD now 15 years.
So when I started at AMD on the back end of the K8 family, through the Bulldozer family, and then, of course, now more recently on the Zen family.
So I couldn't stress it enough that as much as you hear, oh, software is where the world is going, without hardware that software is useless.
And hardware innovation — it's become a lot more challenging, which has made it a lot more interesting and really pushes you to think outside the box going forward.
You want to have the nice, smooth CMOS scaling that we skated through with through the '90s and early 2000s.
For years to come there is going to be enormous challenges in hardware design. All right. You want to take questions?
BORIVOJE NIKOLIC
Yeah, I have a question. This is something that happened. So first, we are doing 4k videos in this class, and these are snippets of 10 minutes or so of a chunk of a lecture.
Because we are delivering most of the stuff remotely and we are pre-recording. So Dan and I got into this.
We are going to do this over a green screen, green screen over PowerPoint, and all of that. We were doing this in Adobe Premiere.
I tried doing that on a three-year-old laptop, and it gave me an estimate of six hours for the first clip, or something like that.
And I got a brand new one, Apple Mac with an AMD Radeon GPU, the fastest one that there is. And then, suddenly that's eighth the time.
These clips that are 10 minutes long are done in 15 minutes or less, spectacular. It changed my life. But then, Adobe pushed a new update last week and it went to 20 minutes.
And you can't keep up with this kind of stuff.
PROFESSOR
It's software bloat, right? It's software bloat somehow.
BORIVOJE NIKOLIC
Apparently it's better and there is many more features, but it is slower. But there was another spectacular thing that happened last night. I started rendering a video.
I started producing this video and the power went out after about one minute. And it finished, it took 30 minutes instead of 20, but it drained my battery.
I had, like, 20% of the battery left. I was like, oh. This is very much in line what Tom was working on.
You can do a lot of things when you're plugged in, but you have to be very careful how these things are doing when they're left on their own on just the battery charge.
What can you say about that? I mean students probably don't know what is happening in the dynamic voltage and frequency scaling. Yet they might have heard about that.
But it, obviously, is a very big deal nowadays.
TOM BURD
AMD had a huge initiative, Mike, from Sam Naffziger. I forget the name of it. MICHAEL T. CLARK: It's 25 by 20.
TOM BURD
25 by 20, that improved performance per watt by a factor of 25 times from 2015 to 2020.
And I think the final calculation, we came in at about, I don't know, 25.6, or something like that.
And absolutely, because of that software bloat, we're continually pushing the envelope on power efficiency.
And we'll continue doing that performance per watt. Because if we come out with a new chip, and yet, you get 12 hours of battery life, 13 hours battery life, you say, hey, that's great.
I bet you in two years of another few software updates you'd be down four hours or less.
So it's a continual battle, but absolutely, you see that in the mobile space in phones, I mean, in all market segments. But yes, absolutely. I think Mark had alluded to that.
I spent a lot of time on the Zen 3 about that to make sure that we could deliver this huge performance uplift. You got to do that while staying within your power budgets.
And there's enormous amount of work and desire that has to go into the CPU to support that, and the system level as well. It's all levels of the hierarchy.
BORIVOJE NIKOLIC
Thanks. Here's some questions. So perhaps this one can go to Mike. What were some of the biggest challenges that had to be solved in order to develop the Zen 3 CPU?
And what were the biggest challenges, roadblocks that AMD needs to overcome to create the next generation of chips? MICHAEL T. CLARK: Yeah, so, Tom was getting to it.
And really, I'm sure you're learning in your classes, too, I like to talk about the wheel of performance. There's IPC, there's frequency, there's area, and there's power.
And a good architect for each design is trying to balance all those. And then you have software on top of it with security and ISO that you have to take into account as well.
So starting with Zen 3, kind of another good property we've learned along the way is that we do a grounds-up design.
Then Zen, with 52% IPC, then we did what we call a derivative — along the way, as architects a lot of stuff changes. We get things wrong.
We see things we want to fix, but because of the schedule they make us put our pencils down. You know how engineers are.
They're always like, man, I just wish I could have fixed that. And so we do a derivative, and then we can get good amounts of performance uplift, 15% IPC on a Zen 2.
But for Zen 3, it's one of those — we've kind of learned we needed to do a whole new grounds-up core again.
If we were really going to stay on that line that Mark showed you of performance and still keep the perf per watt that Tom was talking about, we have to relook at the whole architecture, come up with a grounds-up core, and make sure we effectively use every transistor.
Because as the scaling has slowed down, as Tom was hinting, and costs are going up, we have to be even more prudent and even better architects to make sure every transistor gets used wisely.
BORIVOJE NIKOLIC
All right, thanks. Here is a question that is for Mark, but you guys know this well. How do you see the future of Cisco versus specialized risk processors in the world of IoT?
Will this be a symbiotic relationship? Perhaps not in the world of IoT, if you think about more broadly. MICHAEL T. CLARK: Yeah. I can go first.
Tom can go as well.
Interestingly, even though I've only worked on x86 at AMD for 27 years, to me, ISA is not that big a deal as far as — our machine doesn't directly execute x86 instructions,
right? We break them up into whatever operations we need. We actually fuse x86 instructions together to get more work done per cycle than what was in the actual original instructions.
So we make new CISC instructions in the architecture. We had a couple, of CISC and RISC. So I think it is going to be symbiotic.
And it's really more about the underlying microarchitecture than the actual ISA on top.
So I think if you're a good architect you can design a good CPU in any ISA.
BORIVOJE NIKOLIC
But we told them here in class that the x86 processors are actually RISC data parts. And there is other stuff that makes it appear like CISC. Is that true?
MICHAEL T. CLARK: You can think of it that way. A lot of times you think of it as breaking it down into simpler operations. But again, that's the point.
It's not really the ISA, it's the architecture that delivers the performance. And so I think you'll see. And the design point you're trying to hit, right?
I think that's one of the other challenges we have is we're trying to use this one microarchitecture to go from those supercomputers, down into that laptop, and not make your battery run out in 30 minutes.
So it really is the same basic core across those dimensions.
We do cut down some things and we obviously don't put in as many cores, but the basic root, what we call the engine, is the same in all those devices.
TOM BURD
Yeah, I'm just going to add in that you can go look at architectural.
BORIVOJE NIKOLIC
All right, yeah I — TOM BURD: Oh, go ahead, Bori.
BORIVOJE NIKOLIC
Yeah, it finished, though, in 30 minutes, spectacularly. Other than the GPU, what are some other hardware accelerators are being developed?
MICHAEL T. CLARK: You want to go for that one, Tom, since I — TOM BURD: I don't know to what extent I could answer that, but let's just say we look at the ecosystem around us.
And I don't think that there's any things that we have announced, has there been, Mike?
Probably in AMD, of course, but any high-tech company, there's going to be a host of exploration R and D that goes on in various side programs.
A lot of times those are quite the interesting and challenging things to go off and look at. | could say it's on our radar but I probably can't go into too many more details.
BORIVOJE NIKOLIC
Okay. A couple more questions, and we probably need to let you go soon. What got both of you into the world of computer architecture?
MICHAEL T. CLARK: Yeah, that's interesting. My mom was a teacher and we actually had an Apple II. They had a discount program.
And so we got an Apple II computer when I was in grade school. And I just loved that thing, just playing with it.
But, of course, then more software, but it was just so intriguing to me that I wanted to learn how all the hardware worked inside of it.
And so, yeah, I wanted to be a computer architect as a kid. It's kind of weird.
Yeah. I guess that's it.
BORIVOJE NIKOLIC
Tom?
TOM BURD
I guess for me, there's probably only like one person that might even recognize this, but I remember the ZX80 microcomputer. That's really dating me, I know.
One of the first ones you could buy for home use, and then Commodore 64 after that. I found, like at Berkeley, I think even this class — I would have taken a different class.
But the one that I found most interesting was probably the closest class is Cs 52, and going off and building, at that time, your own Spark CPU core.
And then going off and spending grad school at Berkeley, did the research project building an entire RM8 core, building the infrastructure to go off and do the dynamic voltage scaling.
And that was a phenomenal experience covering all the various disciplines, so that I could take that experience with me, whether it's CAD, whether it's physical design, whether it's RTL architecture.
And the more breadth that you get while you're in school, the higher value you will be in hardware design down the road.
Because if you've already heard Mark hound on that, there's so much cross-team collaboration, that these huge complicated chips are a team of hundreds of people.
And the more that you can talk to people outside your domain in your specific expertise, the more effective you can be to move the whole train forward.
So on a day-to-day basis, me and Mike do pretty different things. But there's enough overlap in terms of our overall long-term roadmap that we've got semi-regular discussions and forums.
And that makes it very interesting and exciting to keep working in this area.
BORIVOJE NIKOLIC
I met Tom when he was a grad student. I had just started. I was junior faculty at [INAUDIBLE].. He was, I think, my first dissertation that I signed up as a second reader.
Bob Ruterson was Tom's advisor. But Tom was a one-man microprocessor show.
He designed with very little help, a microprocessor from scratch, and made it scale its performance with supply voltage, which was the first one of the time.
Quickly, just why does AMD use partners like TSMC to manufacture chips instead of making them in-house? I think AMD went through a painful experience of doing that.
TOM BURD
You wanna take that one, Mike, or you want me to? MICHAEL T. CLARK: No, I can. Yeah. Yeah, fabs are really, really expensive.
And while we did have a CEO who used to say, real men have fabs, he doesn't say that anymore.
But yeah, you really either — because they keep costing so much, either need to continue to develop other types of chips to keep filling them and sucking in more things to fill that fab, that it's a pretty tough business model.
Whereas using TSMC and others, who can take multiple customers to leverage the cost of that fabrication facility is a better economic model for us.
BORIVOJE NIKOLIC
All right. One final question then we'll let you go.
What role has grad school played in your career development, and what are some thoughts about doing grad school right the way after undergrad, versus doing it from industry, or after spending a few years in the industry?
CLARK: Yeah, I can go first. Yeah, to be honest, I was broke.
And I always wanted to go to grad school, but I really needed to get a real job and start paying down my student loan debt. So yeah, but then it was a wonderful opportunity.
At AMD here they had a program where they would pay for you to go back to school. And so I put in the effort to put in nights and weekends.
And we have flexible schedules here to be able to take just one class, spread it out over six years, one class a semester.
But no, I believe in continual learning, the collaboration you get in that environment.
It was 2003 when I graduated, I took a machine learning class, if you can believe it.
And so new ideas come from — there are old ideas that get better, right, as well as just totally new ideas.
I think even if you go to grad school, you need to continue to keep — this industry moves so fast, you just have to continue to keep trying to grow your learning and knowledge.
BORIVOJE NIKOLIC
Tom, how about you? You'd just keep going?
TOM BURD
Yeah, I would say keep going. But I would advise doing internships to get some experience in the industry to help you understand what you'd want to focus on in grad school.
I actually knew somebody we work with, went off for his PhD, and he was probably in his mid-40s.
And I don't know if I could mentally handle that, trying to do that in my mid-40s.
So I think if you're going to go with your PhD, I would recommend doing that sooner rather than later.
Master's, perhaps a little more flexible, but definitely I would highly advise to go do the internships. And that's to help you get some exposure to the industry, absolutely.
BORIVOJE NIKOLIC
Speaking of that, do you guys hire interns? Can you send me a collection of links if you do? Do you hire people with bachelor's degrees? I mean, we have a bunch of those.
TOM BURD
We do. I'll have to send you some links after so you could post for the class.
BORIVOJE NIKOLIC
That would be spectacular. Yeah, we have a few. And somebody just said, yeah, we have 1,200 people.
PROFESSOR
1,200 people, I'm sure there's a couple of them there who'd love to have an internship with you.
BORIVOJE NIKOLIC
Hey, thank you very much.
PROFESSOR
Thank you so much, folks. Outstanding, wonderful. Wonderful to have you here. MICHAEL T. CLARK: These were great questions, and it was fun to be here, Mark.
PROFESSOR
Woohoo, let's give a hand, folks. Thank you.
TOM BURD
Glad we could help out.
PROFESSOR
That's great. MICHAEL T. CLARK: Mark did a great job. His presentation, as always, was great.
PROFESSOR
That was outstanding. That was very good. That was great.
TOM BURD
All right, so hardware.
PROFESSOR
Thanks, folks.
BORIVOJE NIKOLIC
Bye.
TOM BURD
It's the way to go.
PROFESSOR
All right, take care. Take care. All right. All right, folks, let's continue on, have some fun.
I've got the next slide, which is — I'll do a quickie. Let's see if we can do this one-minute slide, Computing in the News. Robot cars without drivers are coming to San Francisco.
I think we introduced this that it was coming to Arizona before, and they're now coming to San Francisco. This is very exciting.
They're going to start neighborhood by neighborhood and try to socialize the idea that cars are going to be coming around without a driver. Don't freak out.
The problem is, what if they socialize in this neighborhood but the person who gets in the car wants to go to that neighborhood, and they haven't socialized yet?
So what are they going to do, not drop them off or something? But anyway, we'll see what happens. This is a GM product and we'll see what that works.
But I'm sure this is going to go viral. It'll be the first.
There's going to be somebody who gets into car, says, ah, and then takes a big picture and a movie and that goes viral.
So we'll see how it works, but fingers crossed for safety, and fingers crossed for the success of that. That is the future and it's here now. It's here in our city, very exciting.
All right. I don't know if you can all see this, and if you Zoom into this — maybe I should Zoom mine in a little bit so I can see it better as well.
Many of you are taking the midterm. I was going to pull up the actual numbers. I had this in another window of the actual numbers of people, who were — where they were.
We are in still in midterm season so we're still processing that. Many of you have lost power. So we're going to announce this in the next slide.
We already announced it, but we have certainly extended the exam from Sunday and Monday to be Sunday, Monday, Tuesday, so that's exciting. So just so you know that that's where we are.
We're in the middle of week, or at the beginning of week 10.
Today's lecture is Caches part 3 and Wednesday's is Caches part 4. We also have a lab this week, lab 7. We're going to talk about that. Labs are no longer required for points.
You attend a session to get those points. But you don't have to do the lab. So if you can do this lab this week, great. If you need more time, it's no problem.
You won't lose points for that. Just show up for your check-in session. It's no longer a check off, it's a check in. We have a whole slide talking about that.
Then, on Friday is OS and VM intro, and then next week we start a whole new topic.
Well, it's still related, but it's part of the memory hierarchy — our virtual memory 1 and 2, and IO is on Friday.
So what's still pending, I believe, I showed here, is homework 7 is due 11-9, and project 3B, people are still working on.
So we're going to talk about that, but that's just the schedule so you kind of know where we are. But there's nothing officially new assigned here.
It's just kind of rollover assignments from before, last week's homework 7 and project 3B. All right.
BORIVOJE NIKOLIC
So if you did not notice it, we moved. We pushed it up two days for those.
PROFESSOR
Correct, okay. So let's now pull to the next slide here. This is the update, and I'm going to pull the scores up, or at least the numbers here.
Here we go. I've got the numbers here. Out of, I'd say, basically, 1,200 — because for the quest section 1, 1,190 took it. So we'll just say 1,200 for the sake of it.
So out of 1,200 who took the quest, I've got about 900 who took section 1, 600 who took section 2, 720 who took section 3, 680 who took section 4, and 580 who took section 5.
So roughly, if we're talking about 1,200, about half of you took sections 2, 4, and 5.
A little bit more took section 3, and 900 out of 1,200, so maybe three quarters-ish took section 1. So people are still taking it, but we do know that there are outages.
And we do wish safety for everybody around, and safety for all of us — certainly that we hope no fires spark up because of the high winds. This is a high-wind event.
As they say, it's the strongest wind event all year, so PG&E, in a preponderance of caution, turned power off for many people so that things don't brush into their lines and cause a fire, as has happened before.
So here's the idea — BORIVOJE NIKOLIC: My house is without power. I'm presently in Cory Hall because I have no internet, no cell coverage, and no power.
PROFESSOR
Right, and this is new information that the cell towers — I thought the cell towers were battery powered or some other power so that even if you turned power off you wouldn't lose a cell signal.
That kind of worries me for, what if power goes out during a big earthquake or something to have no cell phone?
I just think of cell phones as something that's automatically always happens, but that's a significant thing for safety.
BORIVOJE NIKOLIC
Every fifth or so cell tower has the battery backup.
PROFESSOR
Okay. So anyway, exam extended to Tuesday for everybody. So this is it for everybody. We'd love you to take it earlier.
So if you still have power, take it now. Don't take it Tuesday. And there's partly a reason for that, and I'll just read this to you.
If you didn't take it by then, you can just — by the way, if you didn't take it by the end of Tuesday, you say, oh, something happened, something happened, something happened — we'll just put you in the pool of midterm retake students.
And that's going to be 11-7, 11-8. That's going to be after the election. So we're making you do anything between now and the election because of so many things that are going on.
We're going to send out a form for you and figure that out.
We had to now scramble and get TA proctoring for people who had the exam on Tuesday, because we didn't plan for a Tuesday exam day.
And we're still working on that schedule and trying to have people pull extra hours, but we're really stretching our staff thin.
We were going to have them all signed up to do this, and now you still have somebody manning the store on Monday. But now they have to man the store on Tuesday also.
Their workload is less on Monday and now we're stretching into two days. We'd rather it just be you work a certain shift and you answer your questions as fast as you can.
Now the questions are going to come much slower because fewer people are taking it on Monday, and now they have to also work double time.
So it's kind of a shame that we have to push our staff to do this, but we're working with you and we're working with them. And they're also dealing with the same thing.
They're also students themselves. So please take it today if you can. If you don't have any power issues, take it today.
Because we don't have a lot of people tomorrow — just letting you know that. And Piazza is going to remain closed for Tuesday, okay? So that's awesome.
So if you finished your exam, you want to work on the thing and ask questions, you could ask a private question.
But we're focusing and prioritizing all the questions for the exam, just so you know, for one more day, just one more day. We can't slip anymore.
Project 3B — we have extended it to 11-1. So this is Monday, 11-1. I think 11-2 is the Tuesday, or is it 11-3? Maybe it's Sunday.
Let me make sure I get the date right. Sunday, 11-1, okay?
That's two extra days, which means we only actually moved — we only actually stole one day for you in terms of Piazza being closed and all this, but we actually gave you two extra days.
So that is able to react to — maybe you need two days of extra. So we're actually giving two days because people may have actually lost power for two days.
Piazza is going to be closed Tuesday, so you can ask questions. This is very important. We're going to open up a Project 3B text channel on Discord.
But TAs are prioritizing midterm monitoring over project support on Tuesday. Because that's really the most important thing. That's much higher stakes.
It's like a huge chunk of your total course points is the exam, a smaller chunk is Project 3B. So we need to focus on that part of it.
You are not allowed to discuss the midterm on Discord. If you do, we will consider that cheating. Maybe somebody is out here looking at Discord to get answers.
We don't want any of that happening. So if you discuss it on Discord, we are going to pull your name. We're sending you off to student conduct. You cannot do this.
You cannot discuss the exam anywhere, period, on any channel, period, until Tuesday at 11:59 PM Pacific Time, okay? That's the rule, so make sure you don't do that.
Don't get in trouble for that. Make sure you don't think, oh, I didn't know. I can tell you, no, you can't do that. That's a bad thing.
All right.
Here is some more miscellaneous policies just to clarify we're running a little bit over, but we're going to clarify these miscellaneous policies and then we'll be able to take any questions after this.
The quest retake is hopefully happening this week. That's a much lower priority. It's 5% of your whole grade, so we kind of haven't prioritized that.
We're really prioritizing making sure we beta test the midterm and launch that. It's worth much more.
So that's what took over all of our circle of inner circle of TAs who normally would do the auto graders for the quest, and also beta testing the midterm.
They were all focusing on the midterm to make sure we delivered, I thought, a very good product. We've heard some initial feedback, which is people were very pleased with the midterm.
What the numbers are showing, by the way, the mean numbers is they're showing — I'll just show the first question. You have an hour. The average is about 23 minutes.
Remember, I'm not telling you — this is not information. The C part, the class cover is two hours you have. The average is about one hour.
We kind of thought that would be all right. So that's good. Midterm 3 is risk assembly. We thought it would be about 30 minutes or 40 minutes.
The average is about 45 minutes. Section 4, SDS, we have an hour. The average is about 25 minutes. And Section 5 is 35 minutes, even though you have an hour.
So people, mean is well below the number. When the mean starts to creep up to the limit, we're like, oh, no. People are having to cut it off before the end.
Actually, we're doing very well with that. So I think the overall if you add up all these times, it ends up being about 2 hours and 20, or something.
So that's not bad. I told you two to three hours, and the mean is showing it to be that. So I'm very, very pleased that the exam is doing well.
We're hoping to get to midterm grading by next week. There's a lot to do and we have to do it right.
And we have to make sure we cover all the cases so we don't have a lot of those iterations of, please, do a regrade of it.
We want to try to get it right the first time. So we're just going to spend a little bit more time doing that. That probably is going to happen by next week.
So we're focusing on the quest retake this week, and we have to rewrite the midterm for the midterm retake also. So we're doing a lot, folks.
This is much more than we used to do. We used to just write an exam and be done, and there was a couple of Gradescopey, small things.
No, there's like four exams for every exam now. So it's a lot harder than it used to be. But hopefully that's working well for folks who are remote and able to do this.
Lab check-in. We want to clarify what a lab check-in is called. Before they were called lab check-offs, where you got points for being there for answering technical questions.
This is just to clarify, this is no new information. This is just telling you what we really meant if you didn't know that. So the lab completion is no longer mandatory.
So you don't have to finish the lab. If you want to, it's out there.
But it's just a resource for you to use, and it's going to be tested so we recommend you do that.
But we're not going to push you on it because we're not going put all these things. So if you need to do it our week, that's fine.
It's not optimal because you probably need them for homeworks and projects, but you can do that and you don't lose any points for that. The check-ins, again, previously known as check-offs, are still mandatory.
You still need to come to check-ins every week. And again, this is driven by you. You drive what you want to talk about.
You have to come, have a time where you have an appointment, and you have to make that time. If you don't, you don't get points for that week.
There's no do that the week after, the week after. No, you have to come that week. If you missed it, you lose points for that one.
And all it is is a chance for you to have a face-to-face time with a staff member.
And what that means is, if you need to just talk, or vent, or cry, or get help, or get a technical question, anything you want, we're here for you.
We're here to support you, and our staff knows about resources to direct you to for either technical or social-emotional support. If you say, you know what? I'm feeling okay and I did the lab.
Ask me lab questions. Then we're ready with those. We're already ready with those lab questions and the check-off questions you do. There's no points for getting them right.
They're just help, and that's kind of a chance to have a little one on one with the staff member, okay?
BORIVOJE NIKOLIC
We got them. We got those.
PROFESSOR
And those are great, so we have them ready. And as much as possible we're going to have the labs have auto grading components to them.
That's, again, just a free thing for you to practice your stuff, not for any points. But if you can auto grade rather than use up a person's time, do that one.
So we'll hopefully have as much as that. And we've moved homework 7. That was there, so now we slid it back. But we got to have it.
At some point it has to stop, because we have an exam at the end and it backs up, obviously. It backs up like a sink drain does.
So homework assignment is now due the same day as homework 8, which means that's 11-9.
So two weeks in the future, 2 and 1/2 weeks in the future we've got some homework to do, and they're out there.
And by the way, don't forget, you can work in teams on your homeworks. So get your teams together and support each other and work together on homeworks. And we love that, that's just wonderful.
Okay, good stuff. I believe you still have to submit them independently. You work together, and whatever you create, they'll be identical.
You and your partner have identical projects, identical homework assignments, but you need to submit them in Gradescope together. There's no way in Gradescope for you to indicate who your partner was.
That's what we're saying. So in Gradescope it's treating you as a one-on-one person, but you can be an identical solution. Just make sure you have your partner's name on your solution.
Phew, I'm done. Bori, anything to add?
BORIVOJE NIKOLIC
Yeah, get going with this homework, so as long as you have time. Because both of them do have a bit of work. They're two-hour homeworks.
PROFESSOR
Yeah, finish 3B first, obviously. Finish your midterm first, then take a day to breathe and rest, then jump in on that Project 3B. Finish that up. You have till Sunday so that's good.
We gave you some more time, hopefully that's okay. Then take a breath, and then jump back in in your homework. So hopefully that's okay.
And then, farther, farther afield just waiting for you, which is going to be fun, and hopefully you enjoy this one — I'm going to give lectures after Bori talks and does his VM and I/O lecture.
My lectures are going to be thread level parallelism, how to make use of those multiple cores, finally going to learn how to do it in this class, exciting.
We'll teach you that, and then you'll have a project 4 to jump on that. So it should be really fun. People really have enjoyed the parallelism. That should be good stuff.
Okay? Phew, all right.
BORIVOJE NIKOLIC
There's a question about appointments. I think they're closed now, but we got to open them. So I'll remind the TA.
PROFESSOR
That's right. Yeah, I think that's right on the appointments. Okay. Good stuff.
BORIVOJE NIKOLIC
Yeah, we've got to open them so people can sign up.
PROFESSOR
Okay, and we've got still just about 58 people here. Anybody else, anybody else want to share some stuff? It's an exciting weekend. We wish everybody safety.
That's the most important thing with all of the — if they say to evacuate, you should be evacuating, folks.
A lot of people in the Berkeley Hills — I'm not sure how many students live in the Berkeley Hills, but please listen to them when they say to evacuate.
Or even if they're voluntary, not mandatory, please think about that for safety reasons. Because that could be scary. The Berkeley Hills have had major, major fires in the past.
And if you live up in the Hills — I don't know how many people are renting houses up the Hills but I used to live up there.
Bori, I ever tell you I live across the street from where the Oakland Fire was — across the street from the Oakland Fire, and it was a terrible loss of life.
So please don't do that. And I understand that Irvine is on fire right now, and I'm so sorry that folks are suffering that. But I think that that's a challenge.
We had fires up in Napa recently, in Sonoma, and please do reach out if you're having troubles with this. We certainly have extensions.
And by the way, just to remind you, we have an incomplete for anybody who needs it. And I will give you an update. I just want to give you some numbers update.
I see that eight students have taken advantage of that. And that's eight students who are now breathing a little bit easier, now realizing that's off their shoulders.
If you need it we're here for you. So grab that incomplete. It's a pre-approved incomplete. I know of no other classes giving this.
We are giving it to support you. If you need it, you guys have that support, okay?
BORIVOJE NIKOLIC
And don't feel bad about that. You're — PROFESSOR: Yeah, don't feel guilty. It's not like, oh, I'll do it. I'll feel guilty, like we're looking down on you.
That's no problem. If you have it just pull the ripcord and we got it there for you. No problem, folks. We're there to support you.
BORIVOJE NIKOLIC
[INAUDIBLE] next semester. That's fine. one semester at a time.
PROFESSOR
Yeah, no big deal. No big deal. Good stuff.
BORIVOJE NIKOLIC
Okay.
PROFESSOR
All right, I think we're good. I'm not seeing any questions.
BORIVOJE NIKOLIC
Hopefully now I'll get power at some point. I need to record the next four lectures.
PROFESSOR
Bori's going to be in his Cory office with his books. Hi, I'm in my office. I'm recording a lecture, virtual memory, enjoy it. Virtual office.
BORIVOJE NIKOLIC
[INAUDIBLE] I got to tell you.
PROFESSOR
And by the way, thank you so much for Bori for inviting those guests. I'm going to give Bori credit for this.
He's been mostly the one handling and making the relationships with all the guests, so I want to thank Bori for doing a great job with our reaching out to so many.
By the way, this has never happened before and this is a great gift to the class, so I want to thank you what you've done for the students.
BORIVOJE NIKOLIC
We try to get these guys to stay and pay attention to us as well. I mean, you guys are working hard.
I do want to make sure that you're connected to jobs and internships and stuff like that. And AMD is supposedly doing really well. We'll find out tomorrow how well they are doing.
PROFESSOR
Is it earnings report? Yeah.
BORIVOJE NIKOLIC
Maybe that's what Papermaster is working on.
PROFESSOR
I'm sure.
BORIVOJE NIKOLIC
We're only last minute people. I guess that's [INAUDIBLE] as well.
PROFESSOR
But look, Bori, I've been around Berkeley since 1990, that's 30 years. I have never seen an undergraduate course that had this powerhouse array of speakers in the history.
30 years, I've never seen it happen. I don't know about you, too, never.
BORIVOJE NIKOLIC
Yeah. That's the idea. This semester nominally sucks. It's online.
PROFESSOR
Right. At the very least grab somebody.
BORIVOJE NIKOLIC
You can bring people who — there is no way that you know, that there will be people like Victoria Coleman or Mark Papermaster, or [INAUDIBLE].
PROFESSOR
Right, no way. Yeah, this is good stuff. And DARPA — come on, don't forget DARPA, folks. It's a DARPA project, DARPA lead here.
All right, folks. Good to see you all. Stay safe, everybody. Do your midterm today if you can.
Don't push us tomorrow because we don't have a lot of staff to support this, so do a midterm today. Good skill to all of you. Knock it out of the park.
Hopefully you guys are going to enjoy that. Take care, folks. Thanks again. Bye.
BORIVOJE NIKOLIC
Okay, bye. Bye.
PROFESSOR
Yay, thanks, Franklin. Good stuff.
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.