Skip to main content

Optimize Your Prover

info

We are actively seeking partnerships with teams interested in building competitive provers on the prover network. Learn more here.

If you want to be a competitive prover on the network, you'll need to do a lot more than just run the reference implementation. Here's what you need to know:

Multi-GPU Parallelization​

The reference implementation doesn't use multiple GPUs, which means you won't hit the SDK's latency targets. Competitive provers regularly utilize 20-40 GPUs per proving request to maximize throughput. This requires:

  • Writing custom code to split proving work across multiple GPUs
  • Coordinating the RISC-V executor across the GPU cluster
  • Creating checkpoints of the virtual machine state
  • Distributing execution artifacts to different GPUs for parallel proving
  • Orchestrating the recursion prover to combine sharded proofs into a single final proof

Cost-Effective GPU Infrastructure​

Running on AWS or GCP will probably eat up all your profits. You'll need to get creative:

  • Build autoscaling infrastructure for on-demand GPU usage
  • Source long-term GPU reservations from alternative cloud providers
  • Buy GPUs in bulk directly from suppliers

Software and Hardware Optimizations​

The prover is primarily bottlenecked by three operations: hashing, field operations, and memory access. You'll need to optimize at both the software and hardware levels:

Software Optimizations​

  • Optimize assembly of field operations and hashing, commonly used primitives in prover
  • Use AVX256/512/NEON to vectorize certain operations
  • Optimize memory usage and memory access to minimize cache misses

GPU Optimizations​

  • Identify massively parallel parts of the prover, such as merkle tree root calculation, quotient calculation and implement CUDA kernels for them
  • Optimize memory transfer between CPU and GPU

Hardware Optimizations​

  • Source cheap CPU/GPU hardware that is most cost effective and efficient on latency
  • Use more advanced hardware such as FPGA/ASICs to accelerate the prover

Reliability and Performance Monitoring​

You'll need to invest in robust monitoring and testing to ensure your prover is running at maximum capacity and reliability.

  • Robustly test your proving node to ensure you miss deadlines on proofs as rarely as possible
  • Integrate extensive monitoring, metrics collection, healthchecks, and paging to ensure your provers are running at maximum capacity and reliability
  • Benchmark your performance against other provers on the network