Ethereum’s 12-GPU proving problem just got a 4-GPU answer

ZisK's new four-GPU benchmark claim has lowered the headline hardware count in Ethereum's real-time proving race. In February, CryptoSlate examined a roughly 12-GPU setup as a possible centralization risk. The new figure could bring proof generation closer to independent operators, provided the workload and operating conditions are comparable.

ZisK is an open-source zero-knowledge virtual machine project, and on Aug. 18 it said its v1.1.0-alpha prover recorded a 9.62-second p99 on four RTX 5090 GPUs. A follow-up said 99.7% of tested Ethereum blocks finished in less than 10 seconds.

Jordi Baylina amplified the result as a milestone combining four-GPU proving with claimed 128-bit security and post-quantum resistance.

Ethereum's working standard calls for at least 99% of mainnet blocks to be proved within 10 seconds. If ZisK's p99 was calculated over an equivalent mainnet-block workload and timing boundary, 9.62 seconds would sit 0.38 seconds below that threshold.

ZisK has yet to publish the block range, sample size, p99 calculation, timing boundary, proof size, or measured whole-system power for the four-GPU run. That announcement is, as of now, a benchmark claim.

The six-part test

The Ethereum Foundation paired its latency target with five other conditions intended to preserve decentralization: on-premises equipment costing no more than $100,000, power use no higher than 10 kilowatts, fully open-source code, at least 128-bit security and proofs no larger than 300 KiB without trusted setups.

CryptoSlate has also tracked the Foundation's emphasis on the 128-bit security threshold.

ZisK's repository is licensed under MIT or Apache 2.0, and the project claims 128-bit security for the new result. The latency condition remains provisional because the announcement lacks the workload and measurement details needed for an equivalent test.

The hardware arithmetic looks promising while leaving household practicality unresolved. Nvidia specifies 575 watts of total graphics power for a reference RTX 5090. Four cards carry a combined GPU-only rating of 2.3 kW, below the 10 kW ceiling before CPUs, memory, storage, conversion losses and cooling enter the calculation.

Nvidia launched the RTX 5090 at $1,999, putting four cards at $7,996 at launch MSRP. That figure leaves substantial room below $100,000 for the rest of a multi-GPU machine.

A rig can fit beneath a nominal budget and still strain a smaller operator's power delivery, cooling, or host-hardware capacity. Fast proof generation also has to satisfy Ethereum's proof-size and setup rules.

OpenVM shows why benchmark boundaries matter

OpenVM's July production release reported a 9.8-second p99 on eight 5090 GPUs across 7,200 Ethereum mainnet blocks beginning at block 24,000,000. The run used 100-bit provable security and produced proofs under 300 kB.

ZisK's four-GPU number may represent a genuine efficiency gain, while the missing equivalent inputs prevent a reliable ranking. The security levels differ, ZisK has not attached a proof size to its result, and its tested block population and timing boundary remain unpublished.

The Ethproofs API defines proving time as including witness generation while excluding data fetching and proof-submission latency. A p99 calculated around a different interval can look similar while measuring a different operational burden.

Ethproofs currently lists ZisK versions through v0.18.0 and exposes a 16-GPU RTX 5090 ZisK configuration. It has yet to show an independent four-GPU v1.1.0-alpha p99 result, leaving the announcement without a public reproducibility baseline.

ZisK's release history identifies v1.1.0-alpha, while its repository describes the current line as a foundation for a production release undergoing security and correctness audits.

An OpenZeppelin review published in November 2025 examined a limited set of ZisK binary and main constraints at a historical commit. It reported 13 findings, including one critical and two high-severity issues, with none marked resolved in that report.

CryptoSlate's February analysis framed a roughly 12-GPU, seven-second setup as a new centralization risk. ZisK now places a claimed four-GPU run beside that earlier concern, though the configurations have not been shown to be equivalent.

A disclosed mainnet block set, full timing definition, sub-300-KiB proofs without a trusted setup, measured wall power, and an independently runnable release would turn the result into evidence for a home-proving breakthrough.

For now, the claim has weakened the simple data-center-scale objection while leaving the operational case open.