ZKsync's Airbender Prime proves Ethereum mainnet blocks in under 12s on a single GPU

ZKsync announced on October 13 that its next-generation Airbender Prime prover kept pace with Ethereum mainnet blocks on a single GPU, publishing the result on the project's own blog (Introducing Airbender Prime). On October 15, Vitalik Buterin said in response that for zkEVMs the "battle for speed" is "definitively won" and that formal verification and bounded prover cost are now the open questions, per Crypto Briefing.

The numbers

Per the ZKsync blog post, which is the primary source for every benchmark figure below:

• Hardware: single NVIDIA RTX PRO 6000 Blackwell GPU, with additional 2× and 4× configurations also tested.
• Workload: 7,200 consecutive Ethereum mainnet blocks proved end-to-end, including witness preparation.
• Latency: 99% of blocks proved in under 10 seconds. Every block completed in under 12 seconds — inside Ethereum's 12-second slot time.
• Scheme: hash-based STARKs with GKR sumcheck, WHIR polynomial commitments and BLAKE2s hashing. Final on-chain verification uses Keccak-256-based commitments rather than BN254 pairings.
• Stack: RISC-V execution.

The post also publishes a standalone verifier reference, and the implementation sits in github.com/matter-labs/zksync-airbender. There is no engineer named as author.

What's new vs. previous Airbender

Earlier Airbender runs required higher-end data-center hardware (H100-class) to hit comparable throughput. Prime is the first version ZKsync lists at Ethereum-slot-time on a single workstation-class GPU. The write-up does not frame Prime as a sequel to the earlier "Atlas" zkSync OS upgrade; secondary coverage that did was an editorial addition and is not supported by the primary.

Independent framing worth flagging:

• Vendor-reported benchmarks. The 99%/<10s and 100%/<12s figures are ZKsync's own evaluation on a specific GPU over a specific block window. They are not disputed, but they are not independently reproduced.
• Public leaderboard snapshot differs. The live ethproofs.org leaderboard, which tracks "proof weeks eligible" with a 70% target proofs-under-10-seconds criterion, currently does not list "Airbender Prime" as a distinct entry. Earlier Airbender variants sit at 0% weeks eligible on the public tracker today; ZisK (4× 5090) and Axiom (16× 5090) are the top-ranked eligible entries. Readers should treat ZKsync's "leading" claim as vendor-reported until Prime posts eligible weeks on the public leaderboard.

Buterin's response, in context

The Crypto Briefing article paraphrases Vitalik's October 15 statement. The fuller phrasing circulated in secondary coverage is: "At this point, I consider the battle for speed to be definitively won, and the top priority for zkEVMs should shift fully to security." He laid out three near-term priorities for zkEVM teams:

1. Arithmetization-to-spec equivalence — formal verification that the prover's circuit matches the Ethereum execution-layer spec.
2. A hard cap on the ratio of verifier gas cost to the gas cost of the block being verified.
3. Proof-system security, including hash-function and commitment-scheme choices under credible attacker models.

Each one of those is a research program, not a benchmark. The Airbender Prime write-up gestures at (1) and (3) — the Keccak-based final commit is explicitly post-quantum friendlier than BN254 pairings — but does not claim to have shipped formal verification against the Ethereum spec.

What to watch

1. ETHproofs eligibility. Prime needs to accumulate "weeks eligible" on ethproofs.org under the public tracker's own criteria, not ZKsync's. That is the test of a reproducible single-GPU 12-second claim.
2. Formal verification work. Any publication from Matter Labs or an outside team showing circuit-level equivalence between the Airbender arithmetization and the current execution-layer spec. The verifier code and its RISC-V model are the two surfaces that would need audit.
3. Verifier gas cost on L1. Buterin's second priority translates into a concrete number: the Keccak-based L1 verifier call cost per Ethereum block proved. ZKsync has not published that figure yet.
4. Other provers' responses. ZisK, Axiom, Succinct SP1 and RISC Zero are the obvious comparators. Any of them matching Prime's claim on a single workstation GPU, under the public leaderboard criteria, would validate the result as reproducible rather than vendor-specific.

Context

The last year of zkEVM work has turned on moving proving cost from a sustained constraint on rollup economics to a near-rounding-error — the public ETHproofs leaderboard's 70%-under-10-seconds target was set precisely to measure that. Prime is the most aggressive single-GPU claim to date. Whether it survives independent benchmarking will matter less than whether the next milestone — formal verification against the EL spec — ships on the same timeline. Speed is becoming cheap; correctness guarantees against a changing Ethereum spec are not.