Heritage · Still in service

Before video, we made proofs cheap.

From 2022 CFLabs built accelerated verifiable computing: ASIC and software co-design for zero-knowledge proof generation, delivered with partners into shipping hardware. That work is not archived — the silicon is in the field and the research is on this site. It is simply no longer where most of our engineering goes.

11 published papers from that programme, all readable here.

The product

ZKVM + ASIC, as one system.

A zero-knowledge virtual machine proves that a program ran correctly. Doing that honestly costs an enormous amount of arithmetic — which is exactly the kind of workload that stops belonging on a general-purpose processor. Feed a program to a ZKVM ASIC and the proof comes back with more than 20x the throughput of a GPU prover at a fraction of the energy.

Software and hardware co-design

Compute and memory requirements evaluated together, then the system designed around the result — rather than a chip designed first and an algorithm bent to fit it.

Crypto and ZKP IP libraries

Field arithmetic, hashing and proof-system primitives already designed and delivered in tens of millions of shipped chips.

A manufacturing supply chain

Through Intchains Group, a route from design to volume production that a research team on its own does not have.

Case

The Aleo ZK ASIC miner.

Aleo is a layer-1 blockchain that uses zero-knowledge proofs for transaction privacy and scale. Its proof workload is heavy enough that general-purpose hardware sets the ceiling on the whole network. Purpose-built ASIC miners moved that ceiling — we took the resulting hardware apart in public and wrote down how it works.

Read the teardown

Where verifiable computing matters

Three problems it actually solves.

Privacy and scale on chains

Private transactions, and rollups that batch work off-chain and settle with a succinct proof.

Proving a fact without showing it

Demonstrating eligibility, compliance or identity attributes without handing over the underlying records.

Verifiable AI

Proving a model actually ran, and ran on the inputs claimed — optionally without revealing weights or inputs. The line where this heritage meets the video work.

The research is all here.

11 papers on RISC0, SP1, Binius, Jolt, Labrador, Aleo's Varuna, prover networks and field arithmetic — full text, with figures.

Read the papers