The Hardware for the Next Generation of the Internet

Transcript

all right thank you everyone for coming to to this talk um my name is sam cassette um the founder of a company called aligned um aligned does anyone know what that animation is on the right this is conway's game of life uh aligned is fundamentally about aligning the incentives of game theoretic systems right because game theoretic systems allow all web 3 systems to exist um conway's game of life is kind of an example of an epi phenomenal system so uh sort of part of the theme of what i want to talk about today the talk is called hardware for the next generation of web3 and the thesis that underlies this is we mostly think about algorithms right but i want

to ask the question pose the question to you what would you design if you're a protocol designer if you're an engineer in the space what would you design if you weren't limited by hardware if you're not thinking about how you are limited by hardware i want to pose to you that you are actually um so um web3 is all about protocols algorithms data structures this is what we mostly are talking about at this conference right now at scc um so anyone remember e-gold so e-gold was a platform that allowed you to basically do what bitcoin does but it was centralized and it failed for obvious reasons because you can turn it off um hashcash was another invention that was sort of shortly there after that does anyone know what it was originally for

for preventing email spam but this basically invented proof of work right so you combine those things together some merkle trees et cetera you get bitcoin um these are the things that have enabled the industry and let us sit in this room right now um you know fast forward a little bit ethereum took that and you know extended that beyond just those those data structures um and then finally you know zcash right so zcash introduced this idea of zero knowledge proofs um zero knowledge proofs are the first time that we have a different type of mathematics that's being used underneath this right um but these protocols all need work right and when we say work it's kind of a euphemism for actually you need a bunch of hardware doing a bunch of stuff underneath um this is not satoshi but i'll pretend that it is for a second um does anyone

remember who the second person to only bitcoin or mine bitcoin was there yeah so how finney running bitcoin um he was the second person to mine it second person to have a bitcoin get sent to transaction back then they were doing this on cpus right that was fast enough how finney this is probably a 2008 era cpu so i tried to be historically accurate um eventually we moved this up we started using gpus because we can do it a lot faster for that same hashing function just shot 256 hashing but that work again you know it's useless work right the only reason that we're doing that is just to slow down the mining issuance of bitcoin we're not doing it because it actually does something useful right um

fpgas this is actually a butterfly labs fpga if anyone anyone ever tried to buy one of those and they never really shipped them but but you could do it they're a little faster and then finally obviously now we have these rooms full of giant asic machines doing the stuff that was the progression of hardware but it was useless work hardware honestly it really is just for doing that hashing algorithm um so i think there's some other examples of of how hardware has defined what we do in a computing industry right so this is a normal looking stock photo of a data center in normal data centers you know we have servers in them they run things like the early facebook the early aws etc um they have pretty standard architectures just some computers networked together

but when facebook started looking at how do we run multiple like thousand thousands of streams of video that all need to be transcoded into different regions and different formats and fit on 20 different kinds of phones they came up with this thing called the um the f-16 data center design it's a completely different design it has asics inside of it that allow it to do switching it has different control planes in the fabric um this totally changed data center design this is actually an open source kind of thing a lot of people worked on it um it totally changed the way that we think about how do you do video and code transcode stuff like that within data centers at large and this now lets people you know let facebook serve billions of simultaneous users but you had to rethink the hardware in order to

do that right because they had a different kind of workload than you had before so um what i want to talk about is proof of useful work and what we're moving into i think is in the advent of a time in history in the history of this industry when we have computation you can call it work being called just computation um that actually does something useful um it actually does something else so bitcoin just doing shot 256 hashing um ethereum you know also doing hashing hashing you know that's on a gpu or on an fpga most likely in the early early days i don't know if anyone remembers this this is like before it launched people actually talked about maybe using like sedi at home or folding at home uh so we

can search for aliens or we can do like drug research using uh using the ethereum algorithm that kind of you know that kind of went away we just chose to do a normal hashing algorithm but that would have been kind of use of proof of useful work um but uh then you know we have uh zcash zq has used equihash a bunch of useless equihash kind of stuff plus some zkp's which are useful but then now we have this whole new generation of extremely computationally intense things to do for companies like aztec zk sync star queer alio um scroll iron fish as well these are all companies that are doing something that requires something that looks like proof of work it looks like heavy computation but we don't really know what is this layer right what is the hardware that's

going to run on and how do we how do we do that how do we build that stuff um that is one of the things that a line is working on um i think there are two approaches that are happening right now there's sort of a a decentralization first kind of approach um you know companies like alia for instance have built in to the way that they've launched their network a a proof of useful work right the difficulty of the mining algorithm is a zero knowledge proof right so they just said it's just like bitcoin in terms of mining um you can see this is the mining reward 100 alio per block this is in a test net um but they just said let's incentivize people to do a bunch of zero knowledge proofs um and let's have that be the mining algorithm um scroll this

is a tweet from scroll saying you know in addition to the zk evm based execution layer we're building a robust outsourcing mechanism to enable a decentralized network of rollers to generate zkp's they're sort of thinking decentralized first that hasn't really launched yet but it's very very close that's like fundamental and core to design so i said you know decentralization first in their design that's one approach i think it's a really good approach that's one of the things that we're thinking really deeply about if you're thinking about the zk evm for instance um scroll as a kevm that's a lot of workloads right because it's not just taking a tiny tiny proof and trying to roll that through a gpu or something it's actually a much larger piece of data and we need to figure out how to do that then on the other end of

the spectrum you have you know this is the roadmap of starknet at the bottom you can see it says coming soon for decentralized sequencers and provers right so they've taken a let's get it to work first let's build cairo let's have all this stuff obviously they've been extremely successful maybe some people in this room uh know it very intimately and they're right outside um but uh you know aztec zk sync matter labs these companies have taken a similar approach let's let's build something that works and let's figure out later because we know it's centralized right now we know that that doesn't ultimately scale it doesn't ultimately let us get that computation done in a scalable way but we'll figure out later how to build a decentralized proving network but the end result of this is the same thing that happened the

same evolution that happened in this hardware from you know from like cpus gpus fpgas asics my thesis the thesis of aligned is that that's going to happen all over again within a different domain namely in zero-knowledge proofs and namely within other types of high-performance computing that are sort of hiding in the stack of web so how do we actually accomplish this um a few things going on so this is uh this is an example of a few things that are happening right so right now there's something called the z prize if you guys heard of this so the z prize is funded by uh you know polygon amd a few other parties it's basically a competition it's happening

right now about how do we get uh how do we get this computation to be done in a reasonable um equitable sort of you know efficient way that allows people to do it in a decentralized way that allows people to build hardware they're all aiming towards doing something open source which i think is great but if you look at some of the text from the description of the prize we're talking about the fft fft is a fast fourier transform it's a battle about space on the chip doing larger fft on an fpga which is one of the things that we're using starts to require special designs that are largely unexplored except we're exploring them when it comes to more than one card you need to figure out how to coordinate them right so if you think about uh in this bottom right here this is a sort of

a diagram of the the circuits of an nvidia a100 gpu so the a100 is you know basically the flagship high performance computing version of a gpu it's much much larger if you can see on the side on the very very side this is hbm high bandwidth memory so it's the same kind of memory that's in something like a graphics card in the middle are the cores these processing cores the problem is if you have something uh like for instance uh multiscaler multiplications fast fourier transforms etc these things blow up to be much much larger than might fit in a gpu actually this this one's probably big enough to fit a lot of computation in because this is the largest um you know some of these come in like 40 60 gigabyte sizes but if you think about what's the average

ethereum miner using uh what's the average person at home using and even what's the average instant size of something that's in amazon for instance you're talking about like 8 12 16 gigabytes it's too small to fit um the something like something like this a proof size of a large stark right um and so everyone's sort of asking this question and i would contend that the reason why the people that are sort of going to figure it out later like like a you know matter labs or or a starkware et cetera one of the reasons this hasn't been decentralized is because the problem hasn't quite been solved yet of how do you actually do this computation how does it all fit in a way where we can do it at scale and we can do it in a decentralized way these things these a100 gpus they're very very expensive

they really only exist in machine learning context you can't if you even if you wanted to rent a giant farm of them you know an ethereum mine size of them you couldn't uh from amazon or something like that so um so it presents a huge set of problems that the industry needs to face right now so this is this is sort of the point of my talk here is that this is this is what we're aiming at solving um so one of the things you need to think about too is the general disjointness of the different types of proofs that are necessary in order to realize zk rollups in order to realize uh secure multi-party computation these are all things that are you know that are not yet explored not yet understood one of the things we're trying to do too

is standardize this right because if you look across domains if you look across different types of computation that are available then you you end up with everyone doing some something completely different everyone having different proof sizes everyone having different types of hardware that might be necessary for that so when you think about this and what our core competency is at aligned is is that we think deeply about how to stream information through chips we think deeply about what does it mean to have memory clustered in different ways and different configurations if you think about the most advanced teams that are making sort of gpu minor firmware that they ship out maybe it gets you a little bit higher clock speed maybe it gets you a little bit lower voltage or a little bit lower energy consumption

they're not thinking about this at scale how do you take something that doesn't fit in there how do you take 10 of them or 50 of them and hook them together how do you make them then process an arbitrary amount of zkp's that allow us to scale what we all know is is the scaling solution of ethereum which is zk rollups and the industry isn't there yet so we're trying to get there um i think we are you know in many respects there um so one of the responses i usually get when we get our engineering team on the phone with uh another company that's designed something very intelligently about zkp's is someone told us you couldn't do that so i'm this is a very simple and silly illustration of that thing not fitting in that thing i'm sure that uh you know

some of my engineers would find that silly but that's the general idea is that it doesn't it doesn't fit how do how do you get that through um we've come up with other ideas about how to do it like for instance recursive zk proofs that does scale a lot of things but but it also doesn't have the power of being able to compute the entire thing so this is sort of our core competency if you were to go to try to rent something from amazon that not only had a very hefty kind of computational ability but also the ability to access arbitrary memory storage throughput sort of you know super computer level interconnect it just doesn't exist and so that's that's what we're building um so what i want to contend also is this you know this is one type of computation that we all know is really necessary but

there's sort of work aka computation that's just bleeding into the stack everywhere right so block proposer builder separation um this is pretty exciting this is like enshrining flashbots uh or enshrining not the word flashbots um into the next generation ethereum roadmap but we're calling it pbs i think this will engender the creation of uh you know what some people are calling super builders which is essentially validators that exist across different chains are able to do things like uh commute uh compute commitments between chains and do atomic swaps between them this is an emerging kind of field this is a really cool paper by flashbots i don't think it's fully published yet but this is about sort of cross domain mev and what it means to be

a cross domain validator and you know either collusion or being the same validator across those different domains um to do that at scale um you need to be a block builder and to be a block builder what i would contend when i think a lot of people would contend vitalik also mentions this in his end game uh post if you haven't read it you should about becoming sort of like centralization of block builders which is okay it's not necessarily a bad thing um basically being it being a block builder or being a super block builder is a super computing problem right how do you pack the most mev into a particular block how do you compute what mev is going to look like 15 blocks from now this is not something that you can run on a tiny server this is something that

is is a whole new type of computation as well um sort of and beyond that right as as this computation diffuses into the stack we also have things like kzg commitments which are part of dank sharding which i love the name of um so so it's it's dank but it also was made by a guy named um so um so you know uh dank sharding has things that look like you know zero knowledge proofs in it um you know it's hiding everywhere computation is hiding everywhere and we're not thinking about it as an industry in the way that we think and the way that i think we should um so something from spf the other day um he was like you know five to fifty thousand transactions per second is great solana also limited by hardware in many

respects maybe they need a fee market too but um but limited by hardware in many respects uh and but what does it actually take to get to a million transactions per second you're not going to do that on a dell server um so i would contend it involves hardware solutions namely customized hardware solutions in order to make that happen um so it's bleeding through the stack it's sort of everywhere and i want to return to this question if you're a protocol designer or you're you know you're thinking about your application on ethereum or related chains or whatever going zk roll ups um what would you do if you weren't limited by this stuff almost every zkp team i talked to says well we tweaked the algorithm in this way so it would fit in a gpu or it might in the future or something right um what

would you do if that weren't necessary what would you do if thousands of transactions per second hundreds of thousands of transactions per second weren't like that so what a line wants to be by the way is a an anti-thesis to the existing hardware companies in the space which are mostly extractive nothing against bit main butterfly labs in 2010 i think you know said they were going to ship a bunch of stuff and never really did um you know bit main ships these things they they mine on them and then eventually they sell them to the public once they're not mining on them anymore um we want to be a collaborative entity right if you're working on something that you think would benefit from different hardware uh that you know you have a hardware problem that you haven't thought about how to solve or you don't

know how to solve or someone told you you couldn't solve that's sort of our spot um and you know let's think about the stack in a bottom-up and a top-down way i think everyone thinks sort of like you know i have an application and then i have a another layer and another layer and that sort of you know that that software stack is really what most people think about what we think about is actually that at the bottom of the stack it turns into electrons and bits and then it turns into designs and those designs are our domain by the way we also are a full stack validator company we do things up and down the stack but um but our one of our core competencies is to try to help this industry advance by advancing the

science of not being limited by hardware so thank you very much [Applause] is there any questions questions how do you envision data centers validating zero knowledge proofs going forward uh yeah so how do i envision the validation of zero knowledge boost going forward um i imagine a different data

center architecture right so at a line we're building a completely different architecture than you would normally think about it looks a lot more like a super computer highly interconnected ability to share you know workloads across different chips ability to stream information through through entirely custom designs and so i think it looks a lot like that it looks like a high performance computing data center and it looks like being able to serve arbitrary workloads in a way that's not currently done you can't do it on aws you can't do it in a normal like mining farm it looks like a highly interconnected super computing fabric and when do we see that like timelines were starting to see this you know in that part on paper they said

you know this could be bigger than proof of work by orders of magnitude is that next year yeah um so the question was uh when are we seeing this paradigm says it could be bigger than proof of work um i think we're starting to see it right now most of these companies that i had in this sort of roadmap you know this is the end of 2022 middle of 2023 as the they decentralized these approving networks um and then it you know it all depends on transactions per second and the adoption of that i think most of us think that zkp's and zk roll ups are going to get adopted pretty heavily over the next year or so so it's going to scale with that and then there's also a whole multi-party computation so you know the baseline protocol microsoft sap the mpc alliance all these companies are

working on how to use things like zkp's for multi-party computation to replace basically you know the intel sgx chip um and so that's already real we're already really doing this um it's only getting more exponential and bigger so now-ish so yeah thanks um so what do protocols answer when you ask them what would you do if you had unlimited uh space available for ckps uh yeah they usually say well we would have designed the protocol a little bit differently or we wouldn't have spent so much time on recursion or we could add additional sort of like op codes or we

could use larger uh larger sort of functions in what we're doing um and and what we can do can go faster um those are the the domains that people usually think about so sort of faster bigger less work for the developers those kind of domains um yeah so the architecture that you're proposing with like you know the super computer stuff it's you know it's inherently going to be centralized right and you know i guess the stock where people are talking about this you know right now the block production or the prover production the the prover process is centralized but you know at a later stage we'll decentralize it what you're saying is counter to that

like or the architecture that you're proposing yeah so so if we assume that like we are the only prover then of course that's a centralized thing right i mean there's a world in which maybe we become decentralized or we you know we somehow uh effectuate that but the goal is i think of all these networks is to create an open marketplace for proofs right um i hope that we're you know the biggest and monopolistic maybe but but you know as one a business person should but but in general there's no reason why this won't be an open market we want to be one market participant doing this and i think it will spur innovation into the entire industry and people will will compete on it just as they did in other computational demands

but but but you're right i mean if we're the best then you know we might we might win uh and then we have to and okay sounds good thank you