When I scaled up the server last night, I almost placed an order based on the idea: “the more cores, the faster the single proof.” After reading the @Dusk Dusk Prover documentation, I realized that ZK proof generation is a single-thread task. A single proof mainly depends on single-core performance. Adding more workers increases the throughput for concurrent handling, but it doesn’t split one proof across multiple cores to run together.

I think of it like a noodle shop with only one stove per cook. With a strong enough flame, a bowl of noodles comes out quickly; opening more stoves lets them make more bowls at the same time, but it won’t make a bowl in the pot cook any faster. The official minimum baseline is: at least 4 cores at 2GHz+, 8GB RAM, 20GB storage, and a 20Mbps network. For each additional worker, you also need to reserve an extra: 1 core at 2GHz+, 1GB memory, 2GB storage, and 5Mbps bandwidth.

This made me re-check Dusk’s ZK infrastructure. Machine configuration affects both queue length and per-request latency. The Provisioner already includes the Prover, and the roles also can’t be mixed when estimating. If I really deploy it, I’ll first benchmark the single-core proof time, then calculate the number of workers based on the peak request rate, and keep monitoring CPU, memory, and network continuously. Buying a bunch of slow cores may look impressive on the dashboard, but users will still be left anxiously waiting for proofs.

#dusk $DUSK