Benchmark · 2026-11-08
Ethereum mainnet, re-executed on Google Cloud.
2.12Ggas/s
Median of three Google Cloud VMs (c4d, AMD Zen 5) re-executing 30,000 recent Ethereum mainnet blocks: 1.22× revm, the EVM inside reth, measured on the same machines, with the same inputs. Every block's gas used, receipts root and logs bloom is checked against the chain; 0 mismatches in every run.
Method
Same blocks, same state, same machine, both engines.
- Blocks: Ethereum mainnet 25,000,001–25,030,000 (June 2026): 30,000 blocks, 10.55 million transactions, 909 Ggas.
- State: the working set those blocks read, recorded from a verified snapshot and loaded into memory for both engines. No disk I/O in the measurement; the state root is checked when the snapshot is made, not at the end of the run.
- Checks, every block: gas used, receipts root, logs bloom (Eita also re-derives the transactions root and the header hash). A single mismatch fails the run.
- Engines: Eita's interpreter (no JIT), and revm 43.0.3 through alloy-evm 0.39.0, the block executor reth runs, built with jemalloc and asm-keccak as reth builds it.
- Fairness: Eita and revm run back to back on the same spot VM; decoding and sender recovery run on separate threads in both. Cloud hosts vary, so only same-VM ratios compare across runs.
- Throughput: header gas used, summed over the range, divided by wall time.
- What's timed, both engines: execution, the receipts root and the logs bloom, with decoding and sender recovery on decoder threads in the same span. Neither computes a state root or any state trie, and neither writes state anywhere.
- Where it leans toward revm: revm reads state from a plain in-memory hash map, while Eita reads through its snapshot rules; Eita also re-derives the transactions root and the header hash and runs the full header validation, which revm's harness skips; and revm runs without the change tracking reth adds around it (the per-block change sets and reverts it writes to its database). The ratio is, if anything, conservative.
Every run
Three machine families, separate VMs.
| Machine | Eita Mgas/s (runs) | Eita median | revm, same VMs | Eita / revm |
|---|---|---|---|---|
| c4d-standard-16, SMT off (8 cores), AMD EPYC 9B45 (Zen 5) | 2,057 · 2,127 · 2,116 | 2,115.5 | 1,733.5 | 1.22× |
| c4d-standard-8, SMT on (4 cores × 2 threads) | 1,993 · 2,064 | 2,028.5 | 1,737.3 | 1.17× |
| c4a-standard-8, Google Axion (Arm Neoverse V2) | 1,906 · 1,895 · 1,920 | 1,905.9 | 1,496.4 | 1.27× |
Per-block latency, median / 99th percentile, on c4d with SMT off: Eita 12.8 / 45.8 ms, revm 15.4 / 55.0 ms. Eita's peak memory: 11.6–12.9 GB.
Cost
Two cents per teragas.
Google Cloud list prices, us-central1 (C4D: $0.0327 per vCPU-hour and $0.0035 per GB-hour on demand; $0.0138 and $0.0015 spot). With SMT off, all 16 vCPUs are billed.
| Configuration | Engine | Tgas / hour | $ / Tgas on demand | $ / Tgas spot |
|---|---|---|---|---|
| c4d-standard-16, SMT off | Eita | 7.62 | 0.097 | 0.041 |
| c4d-standard-16, SMT off | revm | 6.24 | 0.119 | 0.050 |
| c4d-standard-8, SMT on | Eita | 7.30 | 0.051 | 0.021 |
| c4d-standard-8, SMT on | revm | 6.25 | 0.059 | 0.025 |
What this is, and isn't
- It measures execution: an in-memory working set, not a node reading its state from disk. A production replay from disk is slower, and bound by storage.
- It compares EVM engines (Eita against revm), not full clients: reth adds its database and its own pipeline around revm.
- The engine is open source: github.com/valeryz/eita-evm (MIT). What in it is proved in Lean, and what is checked or trusted: how it’s verified.