General purpose zkVMs provide excellent developer experience in an ecosystem
known for complex tooling. However, they currently require clusters of GPUs to
generate proofs in acceptable time, and are still far away from providing client-side
proofs to the masses, which is required by privacy use cases, for example.

While several optimizations focus on the prover software, compiler-based
optimizations on the VM, ISA and guest program are still underdeveloped.

In this talk we will present our work on these techniques, including automatic
synthesis of ZK precompiles and ZK-friendly program optimization. Our
experiments integrating these optimizers in powdrVM and other zkVMs show that
the techniques are framework- and prover-agnostic, and have the potential to
improve proof generation performance by one order of magnitude and more.