For measurement purposes, we built a custom tool that acts as a proxy in front of each JSON-RPC exposed by Execution Layer clients. This allows us to collect data independently of internal client metrics and accurately compare the performance of each EL.
We ran all clients on the Ethereum mainnet, setting up multiple identical VMs—one for each node—with Lodestar as the Consensus Layer client. We then monitored engine_NewPayload response times for each client.
The tooling for measuring requests: https://github.com/NethermindEth/jrpc-interceptor
We used six identical machines with the following specifications:
| Provider | Akamai (formerly Linode) |
|---|---|
| CPU | AMD EPYC 7713 64-Core Processor (32 vCPUs on single machine acquired) |
| Memory | 64 GB |
| Storage | 1.2 TB (NVMe), ~120k IOPS |
| More details | Premium 64GB by Linode - Public Cloud Reference (cloud-mercato.com) |
| Execution Client | Consensus Client | |
|---|---|---|
| Node 1 | Erigon 3.0.0-alpha2 | Lodestar 1.20.2 |
| Node 2 | Besu 24.7.1 | Lodestar 1.20.2 |
| Node 3 | Geth 1.14.8-unstable | Lodestar 1.20.2 |
| Node 4 | Reth 1.0.3 | Lodestar 1.20.2 |
| Node 5 | Nethermind 1.27.1 | Lodestar 1.20.2 |
| Node 6 | Nethermind 1.28.0 RC | Lodestar 1.20.2 |
Below are the results of the comparison:
To better visualize the results, we performed a percentage comparison using Geth as the baseline (100%), allowing us to see the performance differences between the other clients in percentage terms::