Physical Address

304 North Cardinal St.
Dorchester Center, MA 02124

ZK Zero Knowledge Proofs

Hardware Spurring for Zero Knowledge Proofs


Absolutely Zero Knowledge Proofs cryptography is among the most notable developments in the last fifty years of computer science. Absolutely Zero Understanding Proofs (ZKPs) provide distinct properties that make them essential elements of numerous blockchain scaling and personal privacy services, consisting of ZK rollups like StarkNet, private ZK rollups like Aztec, and Layer 1 chains like Mina, Filecoin & Aleo.

ZKPs are sluggish and costly to produce due to a great deal of expensive math operations. However, with the use of specialized hardware like Field Programmable Gate Selections (FPGAs) and Application Particular Integrated Circuits (ASICs), they can be sped up by 10-1000x.

As users look for more meaningful, performant, and personal calculation, the intricacy of the declarations proven with ZKPs will increase. This will lead to slower proof generation, requiring making use of specialized hardware so that evidence can be produced in a timely way.

The operators of the hardware will require to be compensated for their work, comparable to Bitcoin miners. Ultimately, a complete ZK mining and proving industry will manifest, beginning with enthusiasts creating proofs in their CPUs, then GPUs, then FPGAs. In contrast to Bitcoin, we expect that ASICs might take a very long time to see adoption, if ever.

Why do Absolutely Zero Knowledge Proofs matter?

No Understanding Proofs have two primary use cases.

Outsourced Verifiable Computation

Assume you have some computation that is pricey or difficult to run due to constraints of the platform you are using (e.g. your laptop, a Raspberry Pi, and even Ethereum).

Instead of running the computation on your platform, you should run it on a third-party service that can go back to you the output of the computation rapidly and inexpensively (e.g. an AWS Lambda function, or an oracle service like Chainlink).

Normally, you would require to rely on that the computation has been carried out correctly, allowing the company to output a void outcome, with possibly catastrophic effects.

ZKPs enable a 3rd party service provider to likewise output a proof of computational integrity which ensures the output you received is proper.
Zero Knowledge Proofs
ZK/ Zero Knowledge Proofs
Private Computation

What if you have a calculation that is not pricey to run locally, but you want to hide parts of it? For instance, what if I wish to reveal you that I understand the 1000th Fibonacci number without informing you the number, or convince you I sent a payment without revealing either the amount or my identity?

ZKPs permit you to selectively conceal some or all inputs around a computational statement.

Both of the above use cases have actually manifested in the crypto market in several type aspects (to name a few):.

  • Layer 2 scaling: Verifiable calculation with ZKPs allow L1s to outsource deal processing to off-chain high-performance systems (aka Layer 2s). This makes it possible for blockchain scaling without jeopardizing on security. As an example, StarkWare is constructing a scalable smart contract platform, StarkNet, utilizing a special-purpose virtual device that runs ZK-friendly code. Aztec also enables their Layer 2 programs to run independently, without dripping any info about a user’s deals.
  • Personal L1s: L1 chains like Aleo, Mina, and Zcash permit transactors to conceal senders, receivers, or amounts using ZKPs, either by default (Aleo) or as an opt-in (Mina and Zcash).
  • Decentralized Storage: Filecoin utilizes ZKPs (operating on GPUs) to prove that nodes in the network store data correctly.
  • Blockchain Compression: Mina and Celo utilize ZKPs to compress the blockchain data needed to synchronize to the latest state of the chain into a little evidence.

Provided the above, it is safe to state that as cryptocurrency adoption increases, ZKPs will be required in order to accommodate the increased demand for efficiency and privacy from users, and brand-new kinds of applications and procedures.

ZKPs fundamentally enable scalable and personal payments and clever agreement platforms to prosper but introduce non-trivial overheads which have actually traditionally hindered their adoption.

Why are ZKPs sluggish and how do we make them fast?

Showing a calculation needs very first translating it from a classical program to a ZK-friendly format. This is done either by manually rewording your code to use a low-level library like Arkworks, or by utilizing a Domain Specific Language like Cairo or Circom that assembles down to the required primitives to produce the proof.

More expensive and intricate operations result in longer proof generation times. It is also common that some operations are not ZK-friendly (e.g. the bitwise operations used in SHA or Keccak), resulting in long evidence generation times for what might be an inexpensive operation on a classical computer.

Once your calculation remains in ZK-friendly type, you select some inputs and send it to a proof system. There are numerous proof systems, some named after the authors of their papers (e.g. Groth16, GM17) and others with more imaginative names (PLONK, Spartan, STARK). What they all have in common is that they take a calculation that’s revealed in a ZK-friendly format together with some inputs and output a proof.

Depending upon the evidence system, the evidence generation process may vary, however the traffic jam always winds up being either:

In systems where both FFTs and MSMs exist, about 70% of the time producing a proof is spent on MSMs, and the rest is controlled by FFTs.

Both MSMs and FFTs are slow, but have methods of improving their efficiency:.

  • MSMs are embarrassingly parallel and can be sped up by running them over numerous threads. However, even on numerous cores, if each vector of components is 225 long (that’s 33 million aspects, a conservative complexity ballpark for an application like a zkEVM), the reproductions still end up taking a great deal of time. This suggests regularly duplicating the very same operations, and saturating most of the memory readily available on the gadget. In short, MSMs need a lot of memory and stay slow even when heavily parallelized.
  • FFTs greatly rely on the regular shuffling of information as the algorithm runs. This makes them tough to speed up by dispersing the load across a computing cluster, as shown in DIZK. In addition, they need a lot of bandwidth when operated on hardware. The shuffling indicates that you need to ‘randomly’ load and discharge elements from, for example, a > 100GB dataset on a hardware chip with 16GB of memory or less. While operations on the hardware are extremely quick, the time to load and unload data over the wire winds up slowing operations substantially.

Simply put:

  • MSMs have predictable memory accesses which permits heavy parallelization, but their cost is still high because of the raw amount of compute and memory required.
  • FFTs have random memory gain access to that make them not friendly to hardware, on top of being naturally difficult to operate on dispersed infrastructure.
Zero Knowledge Proofs
Zero Knowledge Proofs

The most promising work we have actually seen on resolving the slowness of big MSMs and FFTs is PipeZK. In their paper, the authors explain a technique to make MSMs more affordable using Pippenger’s algorithm to skip replicate calculation. They likewise explain a method to “unroll” FFTs so they can be carried out without considerable shuffling, which permits speed improvements on hardware due to the now-predictable memory gain access to patterns.

Presuming the above techniques deal with the fundamental traffic jams of each algorithm, the concern then becomes: What is the best hardware to flash with extremely enhanced MSM and FFT algorithms to speed up ZKP generation?

Hardware matters.

The above velocity methods can be implemented on multiple hardware innovations: GPUs, FPGAs, or ASICs. However which one is the best choice?

To answer that, we initially need to acknowledge that ZKPs are still early in their development. There is still little standardization on system specifications (e.g. FFT width or bit-size of elements) or option of evidence system.

Due to these aspects, there are two core homes of FPGAs that make them more effective to ASICs in the ZK context:.

  • ” Compose several times” vs “write once”: The business reasoning on an ASIC is write-once. If any ZKP logic modifications, you need to go back to square one. FPGAs can be re-flashed any amount of times in under 1 second, meaning they can re-use the exact same hardware across several chains with incompatible evidence systems (e.g. since they wish to extract MEV across chains), and nimbly adapt as the ZK “meta” modifications.
  • Healthier supply chain: ASIC style, manufacturing, and implementation usually takes 12 to 18 months or more. On the contrary, FPGA supply chains are healthy, with leading service providers like Xilinx permitting mass retail orders from the website (i.e. with no point of contact) which get here within 16 weeks. This enables an FPGA-centric operation to have a tighter feedback loop on its item, and scale up their operation by purchasing and deploying more FPGAs.

We also expect FPGAs to exceed GPUs for similar reasons that they have actually prospered in machine learning and computer vision:.

  • Hardware Cost: A top-of-class FPGA (leading procedure node, clock speed, energy effectiveness, and memory bandwidth) has to do with 3x more affordable than a top-of-class GPU. The global need for GPUs has actually even more worsened this concern.
  • Power Performance: FPGAs can be > 10x more energy efficient than GPUs, a big factor being the requirement for GPUs to be connected to a host gadget, which is typically consuming a lot of power.

Offered the above, we anticipate that the winning gamers in the market will be business that focus on FPGAs over ASICs or GPUs. If nevertheless only one or few ZK L1s or L2s wind up accomplishing dominant scale, and ZK evidence systems support around a single implementation, then the likelihood of ASICs winning over FPGAs might be higher. However we are most likely several years far from that situation, if it ever occurs at all.


In 2021, Bitcoin miners netted over $15 billion in revenue, and Ethereum miners just surpassed $17 billion. It’s plausible that Zero Knowledge Proofs wind up becoming the de-facto medium of computational integrity and privacy on the internet. In that case, the chance for ZK miners/provers could be of comparable size to the Evidence of Work mining market.

ZKPs are sluggish and will need hardware acceleration to be made feasible over intricate computations. We believe that the technology that will matter the most for ZK hardware acceleration is FPGAs and not GPUs (due to cost and energy effectiveness) or ASICs (due to their inflexibility and long iteration cycles).