NVIDIA cuPQC Cryptography Library
NVIDIA cuPQC is a CUDA-based SDK that accelerates cryptographic primitives on the GPU. Rather than shipping end-to-end protocols, cuPQC exposes cryptographic primitives such as hash functions, Merkle tree operations, and field arithmetic as GPU primitives within a CUDA kernel. Because the primitives live at the CUDA layer, each supported GPU architecture gets its own implementation, expanding support across the full range of modern NVIDIA accelerated computing platforms from consumer RTX cards to datacenter hardware.
Integrating cuPQC into SP1
Proving with SP1 happens primarily through a small set of mathematical operations executed at enormous scale. Our team can only realistically optimize GPU kernels for specific parameter sizes on specific GPUs. Each additional GPU we want to support multiplies that work across different memory hierarchies, core counts, and instruction throughput. cuPQC moves that optimization work out of our codebase and onto NVIDIA accelerated computing, with three key benefits:
- Vendor-maintained optimization. cuPQC remains tuned to all NVIDIA architectures. Even when we change proving parameters or a new GPU generation ships, we get optimized kernels without rewrites.
- Device functions enable kernel fusion. cuPQC's primitives are provided as device functions rather than opaque, standalone kernels. That gives us the flexibility to fuse multiple operations into a single kernel, rather than chaining separate kernel launches with intermediate reads and writes.
- Less bespoke code to maintain. Cryptographic GPU code is exacting to write and expensive to keep correct. Delegating the primitive layer to a maintained, tested vendor library lets our engineers focus on the proof system itself.
Making the Prover Network More Accessible
The Succinct Prover Network (SPN) is a decentralized protocol that coordinates a network of distributed provers to generate zero knowledge proofs at competitive cost for bridges, L2s, exchanges, and other infrastructure providers.
Historically, SP1's kernels were hand-optimized for specific GPUs, which narrowed competitive participation on the SPN to provers who could acquire and operate that hardware. With cuPQC, SP1 will get a consistent interface with optimized performance across the range of NVIDIA GPUs. Operators don't need to match our exact hardware profile to prove efficiently. Teams with heterogeneous clusters, existing datacenter capacity, or newer hardware can bring that capacity to the Succinct Prover Network and compete to generate proofs.
That widens the network in both directions. A larger set of GPUs becomes viable proving hardware, deepening the network's capacity and making proving markets more competitive. And, as new GPU generations ship, cuPQC's optimizations carry SP1 forward with them.
What's next
The cuPQC integration is part of ongoing work to make SP1 the most performant and most accessible zkVM. If you're running GPUs and want to put them to work, learn more about the Succinct Prover Network. To dig into the library itself, see the cuPQC Developer Hub.