Ethereum RPC latency benchmark

Warm localhost performance and cold upstream performance are different products. This benchmark measures them separately and validates every Ethereum response.

Install accelerator Raw JSON proof

1.55ms

Warm finalized-read p50

2.84ms

Warm finalized-read p95

10,076

Sustained RPS in the published concurrency-32 audit

What was measured

Test	Workload	Published result
Warm proof	300 repeated `eth_getBalance` reads at an explicit finalized block	300/300 valid; 1.55ms p50, 2.84ms p95, 3.32ms p99
Sustained load	2,000 validated calls, concurrency 32	10,076 RPS on the measured Mac
Privacy	Foreground edge messages during 300 warm reads	Zero
Correctness	Canonical safe head and JSON-RPC validation	Gate passed; zero RPC errors in proof window

Scope: These figures do not mean an uncached Internet request completes in 1.55ms. A first-ever or expired read still pays local processing, network, Worker, and upstream-node latency.

Run the shipped harness

curl -fsSL https://rpc-relay.pages.dev/install.sh | sh
rpc-relay-pro setup
rpc-relay-pro benchmark
rpc-relay-pro audit

Distribution verified July 22, 2026: the exact downloadable v1.2.0 archive was extracted outside the repository and its included rpc-relay-pro audit command passed. That run returned 300/300 valid warm reads, 2.33ms warm p95, 11,099 sustained warm-cache RPS, and zero added RPC errors. The running accelerator held a non-null multi-upstream canonical safe head; a subsequent two-validation observation window added zero failures.

How validity is checked

The harness rejects HTTP failures, JSON-RPC errors, malformed Ethereum results, missing local-cache headers, warm traffic leakage, and a missing canonical safe head. It reports client wall time separately from accelerator compute time.

Benchmark questions

Why can a cold call take tens or hundreds of milliseconds?

It must reach an upstream Ethereum node. Local caching removes that trip only after a safe result is available.

Does 10,076 RPS mean the Pro edge quota is 10,076 RPS?

No. It is localhost warm-cache capacity in one measured load test. The published Pro service limit is 300 calls per 10 seconds.