We are releasing OpenVM v1.4.0 today under MIT and Apache 2.0 dual-license. This release includes a new GPU prover for OpenVM, new execution system running at 150 MHz, and multi-GPU orchestration for parallel proving on hundreds of GPUs. OpenVM can now verify mainnet Ethereum blocks for under $0.0003 per transaction and in 15s on GPU, an additional 5x+ improvement over the v1.0.0 release in March 2025.
We are excited for developers to continue using OpenVM to tailor ZK for their needs, now with substantial performance improvements across execution and proof generation. By shipping with a fully open-source GPU prover, this release enables developers to run and customize OpenVM however makes the most sense for their use case.Â
OpenVM v1.4.0 is available under MIT / Apache 2.0 license on Github and is recommended for production usage after an audit by Cantina. This release also comes with a substantial update to the developer docs including specs for our new distributed proving design. If you’d like to chat about using or customizing OpenVM for your needs, join our developer Telegram.
Inside OpenVM v1.4.0: GPU Prover and Distributed Proving
The core feature in this v1.4.0 release is a new GPU prover for OpenVM. Developers can now dramatically speed up proving times with GPU acceleration for both proving and trace generation. The new GPU prover supports all Nvidia GPUs with at least 24GB of VRAM, including the L40, RTX 4090, L40S, and RTX 5090 cards widely available for both self-hosted and cloud support. It supports trace generation and proving fully end-to-end on the GPU, with no intermediate host-device communication to minimize latency.
We are releasing the GPU prover under the same MIT / Apache 2.0 dual license as the rest of OpenVM v1.4.0. This enables developers to use GPU acceleration in the most flexible way possible, particularly in different containerized or on-prem hosting environments. We also look forward to seeing customizations and optimizations from the community for different use cases!
In conjunction with GPU acceleration, we’ve shipped a redesign of OpenVM execution and trace generation to facilitate multi-GPU distributed proving. As illustrated above, after a preparatory metered execution, OpenVM now generates traces and proofs fully in parallel, with all non-execution work happening purely on the GPU. This distributed proving support scales to clusters with hundreds of GPUs and can accommodate heterogeneous GPU backends.
Finally, to make distributed proving performant, we’ve substantially optimized OpenVM execution via an optimized interpreter now running at 150 MHz. This incorporates direct memory mapping, tail call optimization, and extension- and host-specific acceleration to dramatically improve performance in a way which will generalize across extensions. In the distributed setting, the new execution dramatically reduces proving latency from the serial metered execution.
A guide to using OpenVM on GPU is included in our updated developer book. To learn more about the architecture underlying the multi-GPU distributed proving and execution, check out the new specs.
Performance Updates
OpenVM v1.4.0 can prove Ethereum mainnet blocks for ~$0.0003 per transaction and in 15s on GPU machines. This is another 5x+ improvement over our March 2025 production release and was enabled by the new OpenVM GPU prover and a complete rewrite of execution and trace generation, most notably to move trace generation entirely onto the GPU.Â
To evaluate performance, we generated STARK proofs for Ethereum mainnet blocks on both a single GPU machine and multiple GPU machines using the new distributed proving architecture. We measured end-to-end proving times as well as proving cost per transaction. All benchmarking was done on L40S GPUs on g6e.2xlarge instances on AWS, and we expect a further 2x performance improvement by switching to 5090 GPUs.Â
On the mainnet blocks we’ve evaluated since the v0.1 and v1.0.0 releases, OpenVM v1.4.0 achieves substantial decreases in cost and latency driven by the new GPU accelerated backend.
We also generated proofs for a 1000 block range of mainnet blocks with numbers between 22,000,000 and 22,000,999 to compare with OpenVM v1.0.0 benchmarks from March 2025. We see similar decreases in proof cost and latency, and also measure the new OpenVM v1.4.0 interpreter running at 155 MHz on this set of blocks.
To evaluate OpenVM performance along different functional directions, we benchmarked a range of tasks including ECDSA and BLS signature verification and iterated SHA2-256 hashing.
These benchmarks are reproducible via our Reth benchmark repo here. As always, we are excited to continue our work to serve more developer use cases by making ZK cheaper and faster.
What’s Next
We are excited to see developers benefit from the cost and latency improvements packed into this release as they deploy OpenVM in production. As part of the open design philosophy behind OpenVM, we look forward to seeing developers run the open-source GPU prover flexibly across hosting setups and customize for their specific GPU or use case. Day 1 hosted proving support is also available through Axiom on the Axiom Proving API.Â
Looking forward, we will continue to push the envelope on performance by innovating on proof systems, zkVM design, and hardware acceleration. We’ve also started work on using OpenVM’s modular design to tailor custom zkVMs for specific use cases – we’ll have more to share here in the coming months. To learn more or collaborate, reach out on Telegram or X.Â
To try out OpenVM today, check out:
OpenVM developer book: A developer guide for using OpenVM to generate ZK proofs for Rust programs and customizing OpenVM for your use case.
OpenVM specifications: Technical specifications for OpenVM, updated with our new distributed proving and refreshed execution framework.
We’re excited to see what you will build with OpenVM – see you on Github!