Scale By The Bay 2019: Bryan Cantrill, Was He Wright All Along? Software After Moore's Law
[Music] all right hey good afternoon I am I'm Bryan Cantrell i if you believe what they say in the program I guess I'm the founder of a stealth company I'm not sure how I feel about that we're not gonna be still for long though so but I'm not gonna talk about that today um what I am going to talk about is is Moore's law and the demise of Moore's law and in particular looking at both the history and and then projecting forward for those of us who develop software and for that I'm gonna go to actually the Talmud here and this is the the original paper that Gordon Moore wrote in electronics magazine and this paper I would really encourage you to read in its in its original form so this paper again in 1965 from electronics magazine and it's kind of amazing to think 1965 that's a long time ago electronics magazine was celebrating their 30th anniversary and they wanted to have a series where the the experts look ahead is what they called it and so they reached out to Gordon Moore then at Fairchild Semiconductor and they asked him to if he would write an article looking ahead on semiconductors in particular and he viewed this is a terrific opportunity because he wanted to make a case that many people in semiconductors believed but people outside of semiconductors didn't necessarily believe and that is that the integrated circuit was going to change absolutely everything about electronics so he he jumped at this opportunity and wrote this this article looking ahead to 1975 from 1965 and this thing is just amazing and I cannot encourage you strongly enough to to read this and indeed we're gonna we're gonna do kind of half a reading here I just I just can't resist it this is the way that the paper actually opens the future of integrated electronics is the future of electronics itself so making that case the advantages of integration will bring about a proliferation of electronics pushing this science into many new areas okay fine then he says this just mind-bending sentence and again the year is 1965 integrated circuits will lead to such wonders as home computers or at least from most connected to a central computer automatic controls for automobiles and just like let's stop there let's not even complete the sentence okay the year is 1965 and he is talking about home computers that is a radical idea in 1965 I think many of us in the room including me were not alive in 1965 and the idea computers were were in the confines of universities and even in a university they would be in the confines of an engineering department or a mathematics department or an applied math department this is before computer science exists as a discipline computer science does not exist as a discipline for at least another decade in most institutions today two decades and he is talking about a home computer like I'm trying to what is the way we would phrase that today it's like a home Space Station it's like what that doesn't even make sense a home computer it automatic controls for automobiles and then you're like wow this is a hell of a set it's like actually I'm not done with my sentence yet also personal portable communications equipment Gordon Moore are you done predicting the entire future in 1965 right this is a stunning stunning sentence and the electronic wristwatch needs only a display to be feasible today so this is the open like big big open here right where you can see and I think it is it is so difficult to see the ramifications of technology in the future and yet he is able to seemingly see it very very clearly I just have to evidence to hide on this because the reason you want to read this in its original is for the accompanying cartoon if nothing else so in this cartoon there's a lot going on in this cartoon you're gonna want to take your time here so the the cartoon and I by the way I've tried to figure out the details of the cartoonist in the history of the cartoon and I can't find any details about it but the cartoonist has cruelly latched on to this obviously insane idea of a home computer and you have this you know Alfred E Neumann figure in the middle selling handy home computers I think the cartoonist is trying to show how ridiculous this is and god only knows what this thing is that he's holding in his hand then you have these two things off to the side you've got the cosmetics booth again I didn't know what this is supposed to be I guess a fare of some sort I don't know what's going on with the woman buy the cosmetics both looks like her back is dislocated and then you have notions over here they don't recognize the word notions cuz I don't wait I'm like what definition of notions is that this is in an archaic a chiefly North American notion chiefly North American definition item used in sewing such as buttons pins and hooks what is the last time that notions was used in a sentence I mean this is published in 1965 I'm gonna guess by like 1967 you could no longer use the word notions casually it seems to be something out of like a sod house this is like this whole idea is is nuts and I think that it shows how crazy the idea was in 1965 of a home computer and again Gordon Moore is gonna make this entire case about integrated electronics and why you know why should people care about integrated electronics and why should people care about the IC why should people care about semiconductors and he's making this case that the principle advantages are going to be lower cost and greatly simplified design as we as we are able to get higher performing integrated circuits and more dense integrated circuits he is a and there's so many great paragraphs in this I love the fact he talks about silicon silicon is likely to remain the base material although others will be of use in specific applications this is at a time when people are looking gallium arsenide they're looking at other conduits other substrates soakin is likely to remain the base material pretty goddamn good prediction in 1965 we are still using silicon yes there have been experiments in other substrates but we still silicon definitely does predominate at not just lower frequencies but high frequencies as well and it still is an abundant and relatively inexpensive starting material so another amazingly good prediction from gordon moore and then he begins to talk about the actual cost of these components and in particular what he is observing is that as we that we develop a new IC and as we we increase the density there's effectively a sweet spot where the the the cost is at its lowest and over time this cost is dropping so and this is actually the first real graph and you know I didn't ask you but maybe we should talk about it what is Moore's law anyway because you wouldn't I mean I think most people probably think rightfully so and this is probably a reasonable inference that Moore's law is about transistor density with the actual the the really the crux of this original paper and the first kind of point he makes is not around transistor density but around transistor economics around the cost per transistor and that cost per transistor over time from 1962 to 65 to 70 you can see that middle of the curve dropping the cost is dropping as the density is rising so to me this is a very interesting graph it's not what we associate Moore's law with we don't associate it as being an economics law we think of as being a density law but he is really making the first point around economics so is that Moore's law well yes and no I mean he doesn't of course make the point that density is rising and that it's going to continue to rise you over the long term the rate of increase is a bit more uncertain although there is no reason to believe it will not remain nearly constant for at least ten years so that means by 1975 again this is 1965 the the number of components per integrated circuit for minimum cost will be 65,000 I believe that such a large circuit can be built on a single wafer which of course was was the case and this is I think what people more commonly associate with Moore's Law transistor density increasing over time now importantly the paper makes no claim as to a law so there is no Moore's law in the paper there's just a bunch of ridiculously and ridiculously accurate prophesizing about the future but there there is no real Moore's law per se the other thing that is that that is is really amazing is where he he talks about some of these other kind of more exotic things that are are out there things like electron-electron repulsion are actually not as likely as federal with and in the near term and then he he says something that I think is just again stunning he says that he points out that shrinking dimensions of the net last sentence shrinking dimensions on an integrated structure makes it possible to operate the structure at higher speed for the same power per unit area what he's saying is that hey there's another piece of good news these things can become cheaper they're gonna become denser and they're going to become faster and this is what's actually known as Dennard scaling but Dennard scaling doesn't get coined for another gate so I don't know if this is actually common knowledge in 1965 or if this is something me this is not something that Gordon Moore talks at all about in his retrospectives of this very important paper but it's a really important observation that we are not only gonna get denser and cheaper but we going to get faster and that's indeed that's what happened right that that is what happened from 1965 to 1975 and then far beyond but the turn wars law is not from that paper the Ford term Moore's law actually was coined by Carver Mead Carver Mead Caltech physicist very incredibly accomplished physicist in his own right and clearly had a missed calling in marketing because he realized that this phenomenon that Moore had identified it needed to be named so he called it Moore's law I love the fact that then Moore I you know cottoned onto this and agreed that this is Moore's law Moore's law was a doubling actually when I grew up going for this way I assumed that it was a doubling of density every 18 months Moore's law from Moore is actually a doubling events the every two years and it was an Intel exec who heard every two years that's basically 18 months right I mean it's like a classic management like wait a minute what like a W density every two years like that's already really good like you didn't need to lie you know what is wrong with you people but it's like at 18 months basically the same thing so but it is that Moore's law as stated by Moore as named by Carver Mead is a doubling of density every two years but it also could be very reasonably inferred to be not just a doubling of density but an increase in speed and an increase in economics decrease in the your their dollars per transistor cents per transistor so that all of those things are our reason we thought of as Moore's law and you know I don't know what your vintage is you know I came up in the in the 1980s and 90s and it really what it was good times for Moore's law because all of those things were happening at the same time density was going up speed was going up and if you're of my vintage you can trace your own lifetime of computing with the clock rates that you had you know these kids today they don't think about their clock rates they'd say you know they don't I don't that you know what are they what are you know my children when they're on a stage in 30 years are they going to talk about like their display resolution or something I don't know whatever it is is gonna be pathetic because what we have to think about is we had you know I was born on a four point seven seven megahertz IBM pc/xt and 8088 and then sure not that long afterward that we had the you know the twelve megahertz eighty my next machine was a 16 megahertz 386 SX I went to college and I had a 25 megahertz spark risk spark which was amazing and that was my freshman year by my by my senior year it was a 40 megahertz ultrasparc by the time I went to my first job I worked at Sun Microsystems for 14 years I well I should have had 167 megahertz ultrasparc but I was like the new guy so I get like the rejects so I actually had 143 megahertz ultrasparc which you actually couldn't buy because the only way to get that was to be the new guy at Sun because the CPU didn't work very well so I have a hundred forty three megahertz by the hundred forty three megahertz and then by 1999 I had a two hundred megahertz Pentium Pro and then we were at a gigahertz and a gigahertz plus and that it was an amazing time right especially I mean I try you try to explain how do you remember the turbo button or the turbo button go explain the turbo button to somebody who has no idea what you're talking about and you will not you will sound insane this is like when you try to explain the you know I I had the the the misfortune of sitting next to a millennial the night of the the 2016 election and she said well at least we gotta know the result tonight and I'm like are you kidding me like don't you remember 2000 and she's like yeah I was like seven in 2000 so not really and I'm like put the whole hanging Chad thing and she's like who's Chad what are you talking about and so I explained hanging chads you ever explained hanging chads someone ever heard of hanging chads she she was like that's you thing okay like nobody has heard of hanging chads I'm like that's not a me thing it's like if you were alive on election night 2008 trust me you know what a hanging Chad is and then she you know a couple weeks later saying hey you know it's funny I got told like you'll ask your parents about hanging chads and she's like I will do it and actually don't I don't care because I don't need the validation I know that everyone knows what a hanging choice and she goes back and she's like hey you know what my parents had heard of a hanging Chad oh my course they were on a comatose anyone does you're the button it's kind of the same thing you go to explain the term a button and you sound nuts so the turbo button the idea of a turbo button was you have a game that's written for a four point seven seven megahertz 8088 and now you're trying to run that on a 12 megahertz eighty or a 16 megahertz three to six SX and the game is playing too quickly Moore's Law is too good Moore's Law has ruined your experience and then in an so like how do you you want a down clock to CPU and you want to make the CPU slower you want to make the CPU act like it's a slower CPU so you can play this dumbass game but in an act of marketing genius they don't call this button which actually is the make the CPU go slower button so no we're not going out guy well how much would you love to know I looked under the actual story I'd love to be in the room where this happened where someone's like all right look the dumb ass State game demographic needs us to slow down the CPU we need a button for this the engineers like are you kidding me like we need a button to make the CPU go slower like what do you call it like the go dumb button what are gonna call that and some marketing genius said no no no no we are gonna call that the turbo button and it will be on by default it's like that's very I see what you done there very clever so that's what the turbo button was the turbo button was just actually the go as fast as the CPU is supposed to go button and when you would depress it you would actually down clock to to at a slower CPU that the turbo button is probably the absolute height of the golden era the era of Moore's law decadence cannot be more concisely expressed than the turbo button and that was great those were good times those were good times but what begin happened very quickly is that the CPU was catching up in terms of its relative speed the CPU was getting faster than any other component that we had and it didn't take long before the CPU caught up so much to DRAM which was not getting any fast was getting denser but no faster and now your CPU was increasingly spending all of its time waiting for loads to be serviced waiting for main memory to actually service the CPU and this is the memory wall and we hit the memory wall by the early 90s it was starting to happen by the late 90s it was rampant and you could have apps that were spending effectively all of their time blocked on main memory we added a lot of caching which was good but it was not sufficient and we're adding caching at a time when the clocks are continuing to go faster and faster and faster so by this era late 1990s early 2000s we are spending effectively all of our time dealing with the memory wall so the the Moore's Law the era of Moore's law decadence is actually over because we've got all these follower on sequential consequences that we have to go deal with and the most obvious one was SMP symmetric multiprocessing adding multiple CPUs that could be processing things in parallel that definitely worked there were other experiments that didn't work very well the pipeline didn't work as well vo iw very long instruction word would definitely didn't work very well if it's if you haven't heard of it you haven't heard of it for a reason as they say those things were basically falling out of favour for single threaded workloads we had this real problem how do you make a single threaded workload how do you make that actually not be blocked on memory and because it's single threaded you don't actually know the result that you need to go fetch until you get the result of the load that it's blocked on well then we had this idea that as it turns out was a very dangerous idea but we didn't realize it at the time which is what's just speculatively execute past the load let's assume that the and what's see how much execution we can actually go do before we need to actually go start retiring instructions in unfolding this and the way that Intel and x86 won and beat out risk in the early 2000s was around absolutely rampant speculative execution that speculative execution had a very dark side that we saw in the beginning of 2018 Mianus horrible Asst 2018 which started with specter meltdown right which were effectively using speculation the speculation in the microprocessor to create side channels whereby you could actually infer what was memory what was it where bits were where bits warned and with surprising efficacy you could actually go and and violate the most basic protections of the machine I'm so expecting of execution had a dark side all of this is to say that let's not romanticize this era too much we can romanticize the turbo button era all we want but by the late 90s early 2000s Moore's law and and dealing with Moore's law was a huge challenge and then Moore's law boy began to end if you ask someone what Moore's law was circa 2000 they would tell you that Moore's law was that doubling of clock frequency and that was a very reasonable thing to say um and it would be kind of like a well actually it's a doubling of transistor density I mean it's like I listen like same thing jerk I mean it's like because it was the same thing for a long time like what you're talking about is Dennard scaling it is true but we didn't need to think about those things separately until the mid 2000s 2005 2006 it's very very clear that we are at the Dennard scaling limit and in particular as we're increasing the clock beyond well beyond regards to gears 3 gigahertz 4 gigahertz beyond we are getting current leakage across the gates and so your efficiency is going way way down your performance is going down it doesn't make sense and indeed the fastest CPUs that we have made from a clock perspective were made in this era we have since reduced the clock quite a bit because it doesn't actually make cents so we it what did make sense is to actually use that transistor density differently so instead of using that transistor density merely for faster clock actually used to have separate course on the die so have separate cores have separate threads on the die and thanks to the work we've done an SMP that was actually a relatively easy transition the problem was that dark silicon really limited what we could go do there and you can't actually have having thousands of cores on a die becomes really really difficult from a power perspective and actually is a variant company called Cerebrus which is actually dealt with a lot of these issues with weight for level silicon but they are very very hard problems to deal with so that's Moore's law and what do you introduce something called rights law because Gordon Moore was not actually the first person to think about the cost of producing things and there's a an engineer named funeral right who it actually took me a while to determine this but I believe he worked at the Curtis Wright Aircraft Company Curtis Wright actually traces its roots back to the Wright brothers and I'm like wayman's it's kinda like related to the right brother he's not I'm no relation I don't think but the funeral Rite was an engineer taking apart the economics of actually building an aircraft and the thing that he observed is that as we build aircraft the costs go down as the volume goes up and it goes down for a bunch of different reasons it goes down because and this is over time so this is not merely an economy of scale there's something else going on and what he observed is we actually better understand how to make this stuff so as we better understand how to make this stuff we spend all of our innovate we spend innovation on the actual process the efficiency the process that drives the economics and the unit cost down and so this is a really interesting law this is a rights law and rights law can be used to explain lots and lots of lots of different technologies and indeed this reinserting group from the Santa Fe Institute led by Justin transecting CO and actually got a bridge the guy named Roy farmer determined that that Wow rights law explains what we've seen with microprocessors slightly better than Moore's law that if we and they try to apply all these different kind of scaling laws and actually writes law could explain microprocessor development up until 2012 a little better than Moore's law which is a really interesting it just should it shows that there can be other underlying phenomenon going on but rights law with respect to microprocessors ended shortly thereafter this is very hard data to get to come across at least without paying handily for it but this guy named Handel Jones who determines the cost per gate over time as we continue to shrink and so as we went from 28 nanometers to 22 nanometers and by the way Carver Mead believed that Moore's law would stop holding at a hundred and fifty nanometers so he believed that 100 150 nanometers we would end we would have the physical limits of Moore's law that's in the 1975 so we are well beyond that with 28 going to 22 but as we're driving the that feature size down it's getting more and more expensive and you can see here from 28 to 22 16 to 14 the actual cost per gate is beginning to rise so we are adhering to Moore's law in the strict sense of transistor density but not adhering to rights law rights law has actually broken down murrs law itself then itself collapsed or is it that we're reaching the end of that Global Foundries pulled out of seven nanometer development that was definitely a big development TSMC and Samsung remain at seven an ammeter and Beyond Intel has struggled themselves to get even to ten nanometer moving beyond this moving beyond seven enemy nor it's the factory where we are today moving to three and five actually requires us to do a big process retooling that's going to be very very expensive now some of that is happening but there are limits to what we can go do and this will end because you may be will be asking like look you know I'm a software person so like forgive the dumb question but like we're talking about nanometers and three anda meters and five nanometers how big is a silicon atom anyway well of course it's actually there are multiple answers to that question but that the because you're dealing at the atomic scale and most of it is empty space but the van der Waals radius of a silicon atom is 0.21 nanometers so when you are at three nanometers that is fifteen silicon atoms across like it's gonna end okay it is absolutely positively gonna end and can we get to three nanometers to five nanometers maybe it will be an at rediculous cost yes and absolutely will be and that's as we begin to look at that that will be the end of Moore's law and Moore's law is not gonna end cleanly Moore's law is gonna end as life ends in confusion and dementia which is what's effectively happening at five DNA age three nanometers it's like you know we should all be so lucky it's just be like hit by a bus and just end it on like those let's make it clean it's not gonna be clean for Moore's law it's gonna be ugly there are deniers it's it's gonna be uh great um no BS we look beyond Moore's law and there are a bunch of things and I'm not time to go into into them in in exhaustive detail there are a bunch of things that are happening but none of these is going to be a salvation specialized computer GPGPU terrific are using our transistor budget to do something else that's great that is going to hit the same physical phenomena and you're gonna hit what what has been called the accelerator wall where you actually can't you're not gonna upend Moore's law I mean you're gonna be stuck on old nodes but even as you get to those new nodes you're gonna hit the accelerator wall um there's you can get away from planner planar transistors and go to 3d 3ds got a lot of Intel in particular Intel has managed to talk themselves into the fact that Moore's law is density per package and that they can maintain Moore's law by going 3d intel desperately wants to maintain Moore's law by the way Intel is like the person in your family that is an absolute denial about the situation and Intel's like I have planned a lovely trip for your mother and me next year and you're like dad instead mom has been dead for three years you and it's like Intel Moore's law is dead okay you've got to come to grips with this um it's like no we're gonna go 3d it's like no you're not gonna go 3d mom's dead you're not gonna go three days heats not gonna work there's a lot of things aren't gonna work here but anyway sorry um 2.5 these a lot more interesting that's what AMD is doing until it's following suit this is where you have with what is called through-silicon vias very strange name but effectively taking different processes within a single dies you've got a single die and laying down different probably allowing a seven nanometer triplet to sit next to a 40 nanometers Chaplin um and this allows us to actually is a much better design that's gonna in this apart the reason that AMD is as kind of lurched ahead of Intel right now Intel will ultimately go that way as well there are alternative physics I love alternative physics I heart alternative physics there are lots of different kinds alternative physics out there i if you get your heart set on one of these that heart will be broken but that's an important part of your maturation as a software engineer is to have your heart broken by alternative physics so I encourage you to fall in love with one of these and get your heart broken mercilessly flash memory being the exception that proves the rule then there's quantum computing quantum computing academically interesting surprisingly real it's also an extremely expensive refrigerator and it's very unlikely that quantum computing is going to deliver economically and economics are what's essential here the idea that I mean if you just look at the refrigeration that's required and they are at I believe the temperature of IBM's the IBM's quantum computer is something like 80 to microkelvin like that's an absolute number it's an it's like 80 to microkelvin is the temperatures like wow it's cold outside yeah it's 80 to microkelvin it's like that again I'm a software guy what do i that seems cold and expensive and it seems very unlikely that you're gonna stick one of those in your pocket again I don't know it just seems like an expensive refrigeration problem I think let's just keep that one let's and it was like but quantum supremacy it's like can we use a different term and not quantum supremacy okay my mom's book but once like calls me from the roof when you say quantum supremacy because she thinks it like the AGI is gonna put everybody out of business okay can we just like keep our pants on about that anyway the quantum computing not that interesting or not immediately interested a Panera pop um is that we are going to see a return to rights law so as we camp out on this seven nanometer node we are gonna see more and more and more compute on the seven Ana meter node maybe two six and five a little bit but we are already seeing this I think we will see this looking five ten years ahead and our return to rights law means that the cost of computer transistor is going to drop over time this massive capital improvement that these foundries have put in to going to a smaller process I think they will put in to greater efficiency incidents the way they fabricate them and that will mean a return to mota to rights law and much more ubiquitous compute so we Assaf for engineers in terms of what are the ramifications for us we are gonna see compute in a lot more places we're gonna see compute at the edge even within the bomb even within the box we're gonna see compute not just in that CPU but on the neck with a smart neck on the spindle and we're gonna see actual like real compute 7 nanometre compute we are gonna see these abstractions everywhere there are a lot of ramifications there are security ramifications there's a lot to be done here and so you know my question is was this actually rights law all along and the economic boom that we have had from what we have called Moore's law and our resulting panic attack by the end of Moore's law maybe this was always rights law and the continuation of rights law is going to continue to deliver what Moore's Law delivered to us all which was more compute more cheaply available solving more problems with that thank you very much [Applause] [Music]