Dr. Christian Killer - Dancing with Moore’s Law: the Mobile Cloud Disruption

Transcript

thank you for the warm introduction so welcome um we're going to talk about acurus today don't worry I will not dance this doesn't work so I'll just click on the laptop maybe somebody can check so as I mentioned before I'm not going to dance don't worry but the title is Dancing with Mor law we're going to look at More's law from different angles and I'm going to outline why this law is important for us uh as a society as an economic system and also for us especially as a uh project uh Gordon Moore is a very famous computer scientist he was a co-founder

of Intel and he said more or less that the number of transistors on a microchip would double approximately every two years while the cost would halven so uh some of you guys probably know about this law and you never think about it but it's actually the defining thing when you look at uh Economic Development around the world because it enabled like everything to scale up from small businesses to the internet itself so it's quite fundamental um looking at it this way especially from uh all these wafer chips silicons that always became smaller and more efficient but we just looked at the performance mainly with Mor law and it sort of hits a boundary right now with uh physics so miniaturizing all these

chips further it becomes harder and harder and we're sort of uh stuck there um right now at least it seems so maybe there will be some some breakthrough there but I think we're um reading reaching some limitations also the industry is sort of agreeing on that so it's not uh me uh perpetrating this this information so what's another perspective to look at things when it comes to um the performance of compute essentially we can look at per performance per watt which means that we don't just look at uh the pure horsepower but we look at um how efficient is it compared to how much

electricity consumption we use uh and this is quite important if we look at the industry um at large we can see that there's a lot of cost pressure everywhere um that's why we have these layoffs maybe um and also there's a lot of pressure um ecologically speaking from from the environment we don't want to use that many resources anymore we want to reuse stuff we want to upcycle stuff and be more um aware of that so Opex as I mentioned is one of these these important things we want to significant we can and we want to significantly decrease them and the ecological aspect so

decarbonization uh also a huge and important topic at large now now um painting our problem that we tackle directly as accurist and also many of the blockchain projects tackle is that we have this issue of 66% uh of the global Cloud Market that is very centralized and if we look at this from a consensus perspective that would be very bad right so we're already afraid of 33% uh with Lio and here we're talking double that so it's quite insane and if you would look at all the subsidiaries and the dependencies that we have of other resellers Etc I think

this number is not even close to what it's really like what's the like what's the effect of this centralization so we have monopolistic pricing we just accepted right if prices go up prices go down we have arbitrary charges we have to check really carefully what's going on in our accounts uh we have a lockin situation um of course from a economic perspective if I'm a um ecosystem by now I want to lock in users so this is what we learn at University this is very good because then we can extract more customer rent right so it's good for a company but for if it's it's like in a monopolistic scale this is a problem so and the

second thing that I'm very passionate about is the data ownership confidentiality so um if you want to run something confidentially you essentially are forced to do setup your own data center uh with your own hsms I talked to a lot of Builders yesterday as well they they do that um because there's no other option because if they run it on AWS or gcp with an HSM there is a lot of trust involved in the provider um and verifying that it's done the way it should be done I mean that's essentially impossible unless you buy the data center yourself which is also possible of course but then you need to be like uh heavily funded so ownership is a problem as well and

confidentiality not just from data address but also data that is being computed on then accessibility um some regions of the world um they are not able to access for example Cloud providers without the credit card um I know there's workarounds but there's always like these arbitragers that uh sell resell for crypto certain cloud in instances but then you have essentially no control over over that instance so that's not really feasible um you have sensorship at play so if um these 66% decide to do a then it will happen so um of course we're also

talking about nation states here for example the US Centric organizations um that really have um a lot to say let's put it this way how can we change that our approach is that we want to leverage mobile devices because um we realized it's a huge untapped opportunity a because mobile Hardware really excels and surpasses um server Hardware on many aspects it's made to be run 24/7 365 um they have highly efficient processors and we know from Chip producers directly that in uh mobile phones they always use the most bleeding

edge technology first so you're always ahead of the servers always of course it depends on what kind of device you use it you cannot just use an 10-year-old Android phone you can for certain computation but not for maybe highly um sensitive stuff and additionally we have these special purpose chips probably you're all aware that you have these tpus inside or you're not tpws and secure elements and secure enclaves or however um the provider calls them outwards and they're really optimized for the compute they execute so this is the opportunity that we realized now I'm taking a step back

and since we're an ethereum conference I guess some of you people realize what this is It's the blockchain trma um popularized by viik don't remember the year um so he sort of said we can optimize for two but optimizing for all three is really hard hence all the other developments right we have uh all these layer twos and C rollups Etc happening but uh Recent research shows so um gladly referencing this later I forgot the reference directly but um we're extending this by adding confidentiality because um lots of problems arise without this

confidentiality um of course the the naysayers are going to say yeah everybody's using it for money laundering no we're talking about just basic privacy needs and confidential Computing needs I mean not everyone needs to see what you're Computing on and the second thing is the effectiveness of that computation not every node needs to execute everything right so this is also an issue not just in ethereum just in the blockchain space at large and the efficiency of the hardware that is being underneath so the infrastructure layer is also not really being looked at uh and hence that centralization in the cloud providers if you go look at certain big projects um that have certain validator structures

everything is run in one data center because it just doesn't work any other way so how can we achieve then permissionless confidentiality which I just mentioned right now um we can use most of these newer devices they have these co-processors they're called so these dedicated chips um s um and we can from within these chips we can bootstrap a um trusted execution environment the trust execution envirment term I'm just using it Loosely here I'm not going to elaborate on the terminology but this is more or less what we do because we leverage the hardware directly and not just virtualization and through that we can

actually have it by Design so we have um physically separated dedicated chips on these devices we can verify it was that chip that executed this code at this point in time and if something is not verifi verifiable anymore we can um just retract and the device will not be used anymore we have no leakage of any data and we have cryptographic designed results so this is like the why um we addressed this from that perspective and now we're looking at some numbers since I claimed now let's put some proof there um as you can see we're comparing a power Edge server to a pixel 7A uh the acquisition cost is like six times lower

for a new device so if you go for an old device it's even less then look at the next number CPU Mark which is like a average benchmarking test for CPUs even the the phone is outperforming the server which is insane and then uh this is my favor part over there the 83 times um is when you look at TPU and GPU performance I think GPU is 50 times and TPU is 83 times so if you have machine learning tasks you can outper outperform a server by 50 to 80 um in terms of factors and then in terms of cost also you can cut cost by 50 times because of that performance per watt that I

mentioned before since we're highlighting more the efficiency um with different architectures there you can save a lot of cost and also energy and you're not using Hardware that was built for a specific server you don't need to uh build new hardware you can even use old Hardware so um yes the G I don't know why it's flowing down there should be up there um we can upcycle devices so old broken devices we can take screens out of broken phones exchange them um as long as the motherboard and the chips work it doesn't even need to work have a working battery unless you want it have a fallback mechanism right and then comparing that we have uh

just $25 um so we have it even cheaper um depending on your supply chain you can get it cheaper uh most likely and also depending on the specific motherboard you look for now we did some evaluations because we wanted to know okay this is all fine and nice but is it really faster so is it really worth it um do all these theoretical calculations also translate to real world performance yes they actually do so we compared Cloud function providers directly um so this was not just a onetime run this were multiple hundreds of runs um this is also under submission right now it will be I can hopefully

publish this in a few days um because of these great double blind uh processes so um we have sorry we have here accurus really outperforming in pure execution time so this is without any latency um of course if you would add a latency you have some more variance in the results but at the end of the day it's still faster so pure execution time of the SE of Tinus so some nerd prime number developers here might know what it is if you ever played with some prime number generation we did this up to 50 million times and we also had to limit it at 50 million because some CL functions would just not return after 75

or 100 million prime numbers so um there's also some limitations there um yeah hopefully we'll put this out soon um what I'm also excited about is that we have true decentralization that's one thing that was talked about at some point um in this space but sort of it got lost always when there's a bull market I believe there's uh people forget about decentralization and everything is going into certain other directions but with um Acuras we can actually achieve true decentralization because we have

resilience in a hetrogeneous mass of devices so we don't have um not even one cloud provider with maybe 10 different server configurations but we have literally thousands of different phones and that's per se a harder attack surface because you hardly can write an exploit for um I don't know anything right you need to have some Target group so this makes it less prone as a single point of failure and theal of service risk we support most Android devices um if you ever developed something on mobile you know how cursed that is so

there's so many Android devices so many configurations and so many like I learn every day of new devices I didn't even know they exist to be honest um on iOS that's a bit simpler but also you're much more constrained so um Android actually was the perfect platform to sort of lock down a device and dedicated to the network whereas in iOS this is harder but we're working on that as well and we have true physical decentralization because we don't have just we're not constrained to Data Centers but we can actually have people building forms or um just leaving them somewhere connected to 5G or starlink or Wi-Fi or even directly to ethernet with

an USBC Port right um so there's a lot of possibility to actually scale this um massively as of now we cover every continent we have around 3,000 devices heartbeating regularly I did a talk one month ago and it was just 2,000 roughly so as you can see the growth is going strong and it's very easy onboarding you scan a QR code APK is being put on the device and you move on with your life going to finish on this note um I mean this statement you can do for anything literally um but um why deepin so more or less deepin came and we were like oh we we realized

damn we are actually a deepen project so um why is it important to me I just want to paint this very clearly because you know and you want to know that bottom up the infrastructure is decentralized and you know the keys are stored safely doesn't really matter what you do at the end of the day that's crucial right um if you don't have that down then there's always some possibility of a detrimental failure to your network Etc all right um just short picture of some Farms so people are fing our test net token um as you can see uh numbers are growing and short call to action join the mobile

Cloud Revolution thanks thank you so much uh Dr Christian killer do we have any questions in the audience hi great do I really lik how you added the fourth dimension to the trilemma so I have two questions um to measure efficiency mean pure physics is somehow output over input so what would you consider the right count like a transaction count in blockchain can be

can mean various things because to claim that there is a higher efficiency you somehow have to have output of transactions in the numerator and kilowatt hours produced consumed in the de Den numerator yeah and second question maybe on decentralization I don't know if you use something like some networks use like an nomoto Co coefficient I didn't get that sorry what what measure would you use to calculate level of decentralization like some networks use something like a nomoto coefficient but my question is rather to you made a point decentralization geographically would

you also not need to be decentralized by counterparty if someone believes in your business which I do and secretly buys lots of secondhand phones and you realize that you have a nice geographical decentralization but it's actually Jeff Bezos who's been buying up all the old phones so most networks you look at Data Center decentralization counterparty decentralization and geographical decentralization I'm just wondering how you would also consider all three wow very very structured and good question thanks so for the first one maybe I didn't make that clear I definitely didn't make it clear so the phones are used not for the consensus layer but for the execution layer so in there you can

for example run V8 JavaScript code and whatever you do in there um is by definition more efficient because he used the mobile device right so that's our efficiency measure I would say for that um in terms of the blockchain layer we use it for the matching of these jobs to the processors which are the devices so we have like a demand site which is developers or people that want to run code and the supply side which are the phones so they provide the execution layer off chain and of course you can do an oracle job on demand in these devices that's sort of the architecture very easily explained I guess this

answers the first question because you were going towards efficiency of blockchain transactions right okay so there's not really a limit there because you just do the matching of the jobs on chain and that works quite well so this and the second thing about decentralization um I completely agree and I think on buying somebody buying a lot of phones per se is not a problem because um the person cannot really tamper with the trust execution environment and that's security and Trust assumption that we have depending on the device so these bleeding edge new for example Google pixel devices um they have huge Buck bounties and so far nobody has managed to to get in of

course that's the Assumption right so um I would say that's the uh the thing and also if you provide um compute to the network you still uh supported the growth of it because um you like you participate in in that permissionless manner so I think that's not a huge threat per se and the last question I don't can you help me with that what was that about counterparty decentralization right I think you answered it so what I would mean with counterparty we run a network actually internet computer and so to be to become a node provider you need to self identify yourself right and

then you also that is the counterparty one and then you need to to certify which data center you will place it in and we obviously know which region you are in but the counterparty is the person physically buying and owning the hardware and you basically want to know that what is the probability of a number X reaching 51% control of your Hardware so that's we mean it's a real person or entity whoever someone needs to buy I mean it's a great case that you make your you make a you make a very strong pitch that we need to rethink the type of Hardware our Hardware costs about 50 $1,000 for one rack MH and so six of them you're $100,000 so exactly it's a real person with real money writing a check exactly um okay to that maybe the

last thing um this is just a protocol and it enables everybody to also make direct deals you can even disable the cost for your processor and I can make offchain deals with you because I can widel list you on the demand side as a process uh like you want a run code Whit list your wallet and you can even pay me off chain you don't even need to use our protocol you can just use the software to directly execute code and this is exactly what projects do they buy phones and then they run oracles on them directly without uh really using um the token mechanism for a permissionless way of course you can do that but then it's

like a public Cloud case right so I think this sort of solves that problem directly for example somebody in um somewhere in Kenya could buy 100 phones and provide like his community with with um compute so thanks we have one more question thank you um you talked a little bit about performance and you said like latency is also one of the things that is important I wanted to ask if you can expand on that and if there any other performance related metrics that you know you thought about or you kind thinking about like now we did um just this prime number generation but we

are now building like a benchmarking that is more testing different things on different servers we also just did it locally just as a benchmark comparison and it's not it's like uh couple hundred milliseconds less than if you run it on a processor so there's this this uh part of the equation and also about the latency um you can also optimize this the larger the network grows um you can choose devices that are closer right so we're also thinking about how to add that in a way that makes sense where is a processor how close is it and it's actually quite easy to evaluate this through just latency measurements right so yeah I don't know if that answers the

question um this is not like a final Benchmark evaluation we're still um going forward there but it's like first step thank you so much thank you so much Dr killer thanks to the audience