Why zkEVM Performance Benchmarks Are Deceiving

Benchmarks measure isolated components. Publicized TPS figures test only the zk-prover in a lab, ignoring the sequencer bottleneck and data availability costs that define real network performance.

The security model dictates speed. A Type 1 zkEVM like Taiko prioritizes Ethereum equivalence, while a Type 4 zkEVM like zkSync Era optimizes for prover speed; comparing their TPS is comparing apples to security oranges.

Real throughput requires real data. A chain's capacity is gated by its data availability layer—whether Ethereum calldata, Celestia, or EigenDA—making prover speed a secondary constraint in most practical scenarios.

Benchmarks measure the wrong thing. Published TPS figures for Scroll, Polygon zkEVM, and zkSync Era measure optimistic execution speed, not the time to generate a validity proof. The prover's computational burden is the real bottleneck, adding minutes of latency that benchmarks omit.

Proof generation is non-parallelizable. Unlike optimistic rollups like Arbitrum or Optimism, which batch transactions efficiently, zkEVM proof circuits have sequential dependencies. This inherent serialization means throwing more hardware at the prover yields diminishing returns, a fact obscured by peak TPS claims.

Hardware costs dominate economics. A zkEVM prover requires expensive, specialized hardware (GPUs/FPGAs) to be viable. The operational cost for Polygon's zkEVM prover network is an order of magnitude higher than a standard Arbitrum Nitro sequencer, making decentralized proving a economic paradox.

Evidence: In production, generating a proof for a 10k-transaction block on Scroll's testnet takes ~10 minutes on high-end hardware. Advertised TPS of 100+ assumes this proof generation happens magically in zero time.

Benchmarks are roadmap signals. Teams optimize for the metrics they publish. A focus on prover time indicates a pre-concensus bottleneck, while L1 verification cost targets post-consensus economic efficiency. This reveals the current engineering triage, not the chain's final form.

The 'full vs. partial' EVM illusion. A high TPS for simple transfers is trivial. The real constraint is opcode compatibility cost. Benchmarks hiding SLOAD or CALL performance signal that full equivalence, the primary value proposition, remains a work-in-progress for networks like Scroll or zkSync Era.

Hardware is the silent variable. A 2-second proof time on a $10k AWS instance is not a network specification. Ignoring prover hardware cost makes benchmarks meaningless for decentralization. This is why Polygon zkEVM and Linea emphasize different aspects of their proving stacks.

Evidence: No major zkEVM advertises sustained TPS under a full, randomized Uniswap v3 workload. The benchmarks that exist use synthetic, compute-light transactions, avoiding the cryptographic overhead of real smart contract execution.