StereumLabs AI

How much machine does an Ethereum execution client need once it sits at the tip of mainnet? Our fleet runs every major execution client on its own host, all with 12 cores, five of six client sets with 16 GiB RAM, one consensus client per pairing on a separate machine. Same chain, same engine-API traffic, one telemetry pipeline. Averaged over the same 36 hours, the synced clients hold between 4.9 and 15.6 GiB of host memory, burn about half a CPU core, and write to disk at anywhere from 0.9 to 22 MB/s. The spread is the story: the client you pick decides whether your box idles or works.

Six Besu nodes, same version, same hardware, same config. Five held a flat JVM heap around 1.0 to 1.3 GB. The sixth climbed about 10 GB a day and was on track to be OOM-killed roughly 30 hours after a restart. The one thing different about it was the consensus client on the other side of the engine API.

This is a walkthrough of how StereumLabs AI, reading our fleet's metrics and logs, took that one anomalous node, traced it to a single method, and filed it upstream. Besu shipped a round of mitigations and closed the issue. A later devnet reproduction showed the underlying layers still pile up, the issue was reopened, and the fix that followed is now in review. The bug is operational: recoverable by a restart, no consensus impact, no double-sign, no state-root divergence. It is also the kind of cross-client interaction a single-node test will never surface, because it only appears when a live pairing lands in a specific state.

A Geth node and an Erigon node sit in the same rack, follow the same Ethereum mainnet, take the same engine-API traffic from their paired consensus clients. Today the Geth host's root filesystem reports ~1,688 GB used. The Erigon host reports ~509 GB. That is a 3.3× spread, on the same chain, with no archive mode involved. Both are full nodes, both synced to mainnet head.

Disk is only half of it. The same pruning that bounds those footprints also runs on the engine-API hot path, and on some consensus-client pairings it pushes engine_newPayload past the 4-second sync-committee deadline. This post covers both halves: how big each client gets and whether you can see it pruning, then what that pruning costs in latency once a node is at the tip.

Key findings:

• 3.3× footprint spread on the same chain (root-FS used per host): Erigon ~509 GB, Ethrex ~598 GB, Nethermind ~1,225 GB, Besu ~1,261 GB, Reth ~1,352 GB (not synced), Geth ~1,688 GB.
• Five of six ECs grew 15 to 19 GB in the 7-day window (~2–3 GB/day, normal block ingestion). The cumulative spread comes from retention policy, not the past week's growth.
• Each EC exposes a different pruning surface, and Ethrex exposes none. Erigon's domain pruner fires ~501 k blocks events per host per week; Nethermind's Hybrid pruner runs continuously at ~1,497 nodes/sec/host; Geth path-mode prunes implicitly; Ethrex has no pruning metric at all.
• Reth is still in initial staged sync: its checkpoint sits ~70 to 130 k blocks behind the tip and is not advancing, so it is excluded from the synced comparisons.
• From Nimbus, three of five synced ECs cross the 4-second deadline on P99 engine_newPayload (Nethermind 4.44 s, Geth 4.67 s, Ethrex ≥5.00 s), but the ranking flips by consensus client. It is a pairing property, not a fixed EC property.
• Nethermind's spikes trace to two deliberate v1.37.1 choices. nethermind_pruning_time spikes past 5 s, over the deadline, entirely from the in-memory pruner.
• Sync-committee duties get hit before attestations. A late sync-committee message misses the next block's sync_aggregate outright; a late attestation still has one slot of grace.

A Prysm node and a Lodestar node on the same chain, on identical hardware, both export beacon_reorgs_total. Over the last 90 days, the Prysm hosts in our Vienna NDC2 fleet incremented that counter 6,011 times. The Lodestar hosts incremented it zero times. Both numbers are correct readings of what each implementation chose to count.

This post is a 90-day reorg census across that fleet plus the smaller GCP comparator cohort: every consensus client (CC) paired with every execution client (EC), against the same Ethereum mainnet, with per-host normalization. The questions we answer: which Prometheus counter to trust for which question, why the same EC behind two different CCs produces very different reorg numbers, and why a "zero reorgs" reading on some clients is silence rather than safety.

Key findings at a glance:

• Prysm increments beacon_reorgs_total 8× more often than Lighthouse over 7 days. The gap shrinks to 1.6× over 90 days.
• Lodestar's beacon_reorgs_total is 0 for the entire 90-day window. Its decline-reason counter fires roughly 54 times per host per week.
• The same EC behind two CCs produces 2–5× different counts: Prysm + Besu reports 69 per host vs Prysm + Nethermind 280.
• Geth is the only EC in our fleet whose Prometheus reorg counter increments at all. Nethermind and Reth export the metric but it never increments; Besu, Erigon, and Ethrex don't export one at all.
• A single fixed beacon_reorgs_total alert threshold does not port between consensus clients. Re-baseline per CC × EC pair.

Ethereum runs on two layers: an execution client (EC) handles the EVM, transactions, and world state, and a consensus client (CC) handles Proof-of-Stake fork choice and validator duties. Six EC implementations and seven CC implementations exist in production today, paired in dozens of combinations across the network. StereumLabs runs all of them, side by side, on identical hardware in our Vienna data center, and publishes the numbers.

This post is the technical introduction to that platform: the bare-metal fleet, the metrics and logs pipeline, the label conventions that make the pairings comparable, and the in-house AI workflow that turns the resulting telemetry into the blog posts you are reading.

RockLogic publishes a separate case study on the business side of this workflow: how the same "AI on own data" pattern keeps customer telemetry inside the perimeter while still producing useful answers. This post is the technical view from the other side of the same workflow.

Before The Merge in September 2022, an Ethereum node was a single Proof-of-Work client. Geth, Parity, Erigon, Nethermind, or Besu, each handling everything from networking and the EVM to mining and chain selection inside one binary. The Merge split that role in two: an execution layer (EL) for the EVM, transactions, mempool, and world state, and a consensus layer (CL) for Proof-of-Stake fork choice, finality, attestations, and block proposals. The two halves talk to each other over the engine API, a JSON-RPC channel authenticated with a JWT secret.

Most operators run those two layers as two separate binaries that talk over the engine API. Caplin v3.3.10 collapses them back into a single Erigon binary. We added a Caplin standalone host to the StereumLabs fleet on April 8, 2026, and let it run alongside the classic Erigon plus CC pairings under identical mainnet conditions. This post reports what we measured: where the monolithic architecture wins, where it pays a tax, and where it leaves observability holes.

Every execution client connects to different peers, sees a different slice of the network, and churns through connections at wildly different rates. We queried Prometheus metrics and Elasticsearch container logs across our entire fleet to find out what each EC actually experiences at the P2P layer, and what it means for Ethereum's network health.

We measured gossip verification, KZG batch verification, data column reconstruction, and getBlobsV2 latency across 72 nodes running all six consensus clients. The verification speed spread between fastest and slowest client: 50x.

We deployed all 6 Ethereum execution clients from scratch on bare metal and tracked exactly how long each one took to fully sync. The results range from impressive to alarming.

How does each execution client behave when Nimbus asks it to build a block? We monitored 1,000 validator pubkeys across 5 EC pairings for 48 hours and found that block building performance varies dramatically, with one client producing near-empty blocks while the other four packed in millions of gas.