Tachyon: A Much Faster and Lighter Halo2

ETH Vietnam 2024 · November 2024 · 11:42

Ryan Kim presents Tachyon, a much faster and lighter implementation of the Halo2 zero-knowledge proving system. The talk covers the performance gains over the standard Halo2 stack and what they mean for developers building ZK applications. Delivered at ETH Vietnam 2024.

Speakers

Topics

From ETH Vietnam 2024, hosted by ETH Vietnam.

Transcript

please join me in welcoming Mr Ryan Kim jkp lead he will diving into the Intriguing topic of how do we optimize hallo to with ton ladies and Gentlemen please give a big Applause to Mr Ryan the stage is yours want to check check hello everyone I'm Ryan Kim and I'm am zkp lead on the chroma team okay today I'm so excited to present the about the latest advancement for the field of CK evm and first we're going to take a quick look at both chroma and takon and dive into Benchmark result in comparison with scores here to

if you heard of it and I'm going to explain the straightforward migration process from her to to takon and then I'm going to explain the some of strategy that we've employed to make this result and last I'm outline the our exciting plans for the future so what exactly is chroma chroma is the first off world with active GK for proof and is essentially is an optimist rup powered by CK evm and this design was chosen for to uh primarily due to the challenges that is

related to the G K EVMS cost and time and we recently achieved historic Milestones that for the first time ever active GK for active challenge took a place uh on our main net of universal world and chma CK evm currently depends on scrs her to for both front end and back end but we are confident in enhancing the back end performance with takon using GPU acceleration while keeping the circuit

part untouched so let's dive into takon it's a modular GK back end excellent by GPU designed to maximize the GPU utilization ratio so let's consider the optimal use of a GPU the fear that immediately come to mind is De learning right and de learning model is often developed using python but internally when leveraging GPU acceleration they rely on C++ so why is that well using zpu means you have to write code with Nvidia SK and we are

only provideed with C++ then let's shift our Focus onto the uh cry blockchain and and uh uh which language do we tend to use more for CK is C++ or Rost of course rust right and Rost is an excellent programming language but when considering implications of using rust with GPU for GK is that it requires rewriting some part of your code in Plus+ again such as MSM or

NTT which results in um which results in some performance drawback because um some of because some of your there's a limit with limit the how much of your code can affect actively run on GPU and this results in performance partnet and because one section of code can effectively run at a time so our vision with takon is to transend these limitations by enabling entire for enabling for entire source code to run simless on a GPU by

ensuring that GPU is free re utilized without any downtime ensuring that performance Improvement across the board and to achieve this goal we initially implemented the her 2 CPU version in ton and guess what we are performed scross here to by 10% and this is particularly not worthy because despite requiring for simple migration additional requirement for simple migration is still surpasses our expectations although it was slower

before Thea phase but takon improveed performance after Theta phase as it gains other the necessary context to generate a proof and moving forward we are fully committed to enhancing the speed of for all the phases it involved and not only have we made takum faster but also we have reduced memory like 30% but simply there are two ways to represent polinomial uh which is one one is the coefficient form and the other is

evaluation form and interestingly we don't have to maintain both firms at every phase rather one firm is required at each phase and however here two maintains both forms for all the polinomial except for advice column so after recognizing this inefficiency we resolve this issue and successfully deduced like approximately 32.625 your circuit at

all typically sequence to the sequence to generate proof follows these steps defining your circuit generating the parameters for the prom Comm commitment scheme and generating The Proven key and creating the random number generator creating the transcript and creating the proof with all the AR arguments above and it's worth taking a note that for takon is that step one is not modified as we want to keep the circuit part untouched takon needs additional context and we can obtain this context by

serializing The Proven key which introduces an an additional overhead which is unique to takon amount to be like tens of gigabyt and subsequently similar structur should be created from the C++ instead of rust and by ensuring that the production of identical proof and here are the optimizations we have done so far and are planning to implement which are lazy fffft bat normalize optimizing paration reducing H hip of allocations optimizing MSM entity

ship approver H poly Builder reducing redundant pomal evaluations creating a pered table just once um those are very documenti in our G repository so if you are curious about this please go please come visit our G repository and we are welcome for you guys as a developer to contribute to our kab to resolve one of these problem list and we takon is also listed in her2 ecosystem like here and thanks to the

overwhelming interest from developers takon once topped a chart on C purpose weekly Top trending and at that time which is above Bitcoin and earning more than 7,700 GitHub stars and we are confident in uh confident in our source Cod becomes one of the best source Cod over entire ecosystem and we believe that our project will be loved by everyone and proves invariable in various fields and this project is our so

already comprised of over 140,000 lines of code and for the next Milestones our plan includes several key objectives we are exploring the possibil possibilities of running her two enti on GPU and we're going to apply foring schemes to ha to and since this is is designed as modular we are fully committed to S supporting as many front end as possible and circum supporting is almost

completed and with kylo is the next targeted front end and lastly we continue to do more optimizations for further enhancement if you want to learn more about takon please scan this QR code and which direct you to our medium article and the links are on the left side is about chroma and the QR code on the right side is a QR code that which which take you to the takong kab repository

and you can re out to me with Twitter or telegram using this contact information below that's all thank you for listening have a nice time in if F now thank you Ryan