We’re building a graph-based zkVM on top of the @ziskvm system. Our goal: rethink prover architecture from first principles to unlock better performance, portability, and security. 🧵

ZKP proving is entering a “speed-first” era. Almost every prover now supports GPU acceleration (CUDA, Metal, etc.). But the architecture behind hardware offloading matters just as much as raw hardware.

There are two main approaches today: - HAL Mode: same function calls, different hardware backends. - Graph Mode: encode the entire proving process as a computational graph, then optimize and compile per platform.

The split is clear: - AI systems → Graph Mode - Most ZKP provers → HAL Mode (for historical, CPU-first reasons) HAL is convenient, but hides the global structure of computation.

Graph Mode makes the entire prover explicit and inspectable, enabling deeper optimization, cross-platform execution, and stronger security & auditability. In upcoming post, we’ll share how this design shapes our graph-based zkVM and what it unlocks. ⚡️