Why Your Rollup's TPS Benchmark Is Probably Wrong

Your TPS is a vanity metric. It measures a synthetic, best-case workload that ignores real-world constraints like mempool dynamics, state growth, and cross-domain messaging.

The benchmark is the bottleneck. A rollup's true capacity is defined by its slowest component—be it the sequencer, data availability layer, or state access patterns, not a theoretical compute limit.

Evidence: Arbitrum Nitro's 2M TPS claim is for fraud proof verification, not user transactions. Real throughput is constrained by Ethereum's calldata costs and the sequencer's mempool processing.

Benchmarks measure optimal conditions. They use simple transfers or identical contract calls, ignoring the computational diversity of real on-chain activity. A network processing 10,000 identical swaps is not equivalent to one processing 10,000 mixed DeFi, NFT, and gaming transactions.

Real load is adversarial and unpredictable. Users submit transactions that maximize their own profit, creating state contention and complex execution paths. This exposes bottlenecks in the sequencer, mempool, and state access patterns that synthetic benchmarks miss entirely.

The testnet-to-mainnet gap proves this. Networks like Arbitrum and Optimism demonstrate stable throughput under controlled loads but face congestion during airdrops or major NFT mints. Their sustained TPS under stress is the only metric that matters.

Evidence: The mempool is the bottleneck. A sequencer claiming 100k TPS in a closed test will choke on a flood of MEV-bundled transactions from Flashbots or Jito. Real throughput is gated by the sequencer's ingestion and ordering logic, not just execution speed.

Benchmarks measure isolated execution. Rollup TPS counts only L2 state updates, ignoring the cross-domain messaging required for finality on Ethereum. A user's transaction is not complete until proven on L1.

The tax is latency, not just gas. Protocols like Across and Stargate optimize for cost, but the sequencer-to-L1 finality delay creates a multi-block confirmation window where funds are locked.

Proof posting is the bottleneck. Even with 100k TPS, a ZK-rollup like zkSync must batch and post a validity proof to Ethereum, which acts as a global throughput governor for all connected chains.

Evidence: Arbitrum Nitro processes ~40k TPS internally, but its Ethereum calldata submission is rate-limited by L1 block space, creating a practical ceiling far lower than advertised.

Synthetic benchmarks ignore state growth. Your 100k TPS claim assumes a clean-slate state. Real applications like Uniswap and Aave create complex state dependencies that bloat the Merkle tree, slowing proof generation and increasing L1 settlement costs.

Data availability is the true bottleneck. High TPS requires cheap data posting. Solutions like Celestia or EigenDA provide cost relief, but they shift the bottleneck to cross-chain bridging latency and security assumptions, creating new trade-offs.

Real throughput requires real users. A benchmark transferring ETH between two funded wallets is meaningless. Realistic workloads involve token approvals, NFT minting, and DEX swaps, which have 5-10x higher gas costs per logical user operation.

Evidence: Arbitrum Nitro's peak of ~40k TPS was achieved in a controlled, single-application stress test. Its sustained, multi-app mainnet throughput is two orders of magnitude lower, proving the benchmark gap.