◆ BENCHMARKS
How we measure. And what we won't fake.
Our edge is physics: co-located bare metal at Equinix FR13, beside the Solana stake clusters and the Jito Block Engine. This page documents exactly how we benchmark - and we publish numbers only once they're measured against a real workload. The first dated reports, raw evidence, and reproduction instructions below.
// flat monthly plans, no credit math · view pricing →
◆ WHAT WE MEASURE
The three numbers that decide a trade.
P50 / P99 latency
Round-trip time from request to response. We report the median (P50) and the tail (P99) - for trading the tail matters more than the average, so we never hide it behind a single number.
Landing rate
The share of sendTransaction calls that land in the target slot window under real load. A fast send that doesn't land is worthless, so we measure landing, not just submission.
First-seen delta
How much earlier the preprocessed stream surfaces a transaction versus waiting for confirmation - measured against your own subscription, not a synthetic feed.
◆ WHERE & HOW
Measured where it actually runs.
Where
Co-located bare metal at Equinix FR13 (Frankfurt), racked beside Solana stake clusters and the Jito Block Engine. Single-region today; the edge is physical distance, not a marketing claim.
Against your workload
We measure with your methods, regions, and traffic shape - not a best-case synthetic loop. Numbers depend on where you are and what you do, so one hero figure would be dishonest.
Window
Rolling 24-hour measurement, P50 and P99 reported side by side, landing rate as a share. Each published report is dated and carries its own methodology so you can check the work.
◆ REPORTS
Published runs.
// latest current-product run · 2026-07-11 · 07:56 UTC · SubscribeDeshred vs processed gRPC →
Deshred first
99.92%
Median lead
4.64 ms
Paired sigs
55,424
Window
5m 18s
Evidence for products available now appears first. The DoubleZero report remains public as network-path research; it is not a claim that DoubleZero-backed fastest shreds ship with a current self-serve plan.
SubscribeDeshred versus processed gRPC
55,424 matched Pump AMM transaction signatures comparing rpc edge pre-execution deshred with normal processed Yellowstone gRPC.
Decision: Assess whether preprocessed data can surface matching transactions ahead of processed gRPC for this workload.
Limit: One Frankfurt Pump AMM signature set, not a guarantee for every filter, region, or transaction.
View report · full data →QUIC transaction relay, the first 1-hour run
2,202 transactions sent through the co-located QUIC relay (Yellowstone TPU, Frankfurt). First-seen (deshred) and processed timing matched by signature.
Decision: Inspect submit-to-first-seen behavior and the full tail before testing your own transaction route.
Limit: A one-hour public baseline, not a provider comparison or service-level guarantee.
View report · full data →DoubleZero IBRL versus public Internet
One-hour leader-paced A/B from Frankfurt: for every eligible Solana leader window (DZ TPU open), one transaction through DoubleZero IBRL and one through the public Internet. Same host, relay binary, transaction shape, blockhash, slot trigger, and leader TPU endpoint.
Decision: Judge where private network routing changed leader-paced placement as leader distance increased.
Limit: DoubleZero IBRL research only. DoubleZero-backed fastest shreds are not included in current self-serve plans.
View report · full data →Looking for continuous operational signals? The solbench live board is pending co-located numbers - separate from these reviewed, published reports.
// want it measured against your own stack? start an onboarding benchmark - DM us on Telegram. last updated 2026-08-13
◆ REPRODUCE IT
Check the work, don't trust it.
Open harnesses
Every report comes from an open-source harness you can run against your own endpoints. solana-tx-bench ↗ · solbench ↗
Raw evidence
Each report links its dated raw output - tails included, nothing trimmed. 2026-07-13 ↗ · 2026-07-11 ↗ · 2026-06-30 ↗
Methodology docs
How each measure is defined, what it can and cannot claim, and how to rerun it. transaction-sender ↗ · stream-methodology ↗
◆ BENCHMARK FAQ