Benchmarks
Published rpc edge stream and transaction-sender benchmarks, with methods, raw evidence, and reproduction instructions.
Low-latency claims are useful only when the lifecycle stage, filter, sample count, host, and tails are explicit. Our benchmark hub contains the reviewed presentation; every report links to its raw evidence and open-source harness.
Published reports
DoubleZero IBRL versus public Internet
On July 13, 2026, a one-hour leader-paced A/B from Frankfurt submitted one transaction through DoubleZero IBRL and one through the public Internet for every eligible Solana leader window (DZ TPU open). Both arms shared host, relay binary, transaction shape, blockhash, slot trigger, and leader TPU endpoint. Across 201 same-slot pairs, DoubleZero was earlier or tied on preceding PoH ticks for 98% of pairs (91 earlier · 106 same · 4 public earlier). The PoH edge scales with leader distance and is negligible within Europe.
SubscribeDeshred versus processed gRPC
On July 11, 2026, we matched 55,424 Pump AMM transaction signatures on the same public rpc edge connection from Frankfurt. SubscribeDeshred arrived first for 99.92% of matches, with a 4.64 ms median lead over processed Yellowstone gRPC.
Public QUIC transaction-sender baseline
The June 30, 2026 initial run sent 2,202 transactions through the co-located QUIC relay. 2,121 (96.32%) had matching landing evidence; submit-to-first-seen was 104.04 ms p50, 374.93 ms p90, and 837.78 ms p99.
Reporting policy
- reviewed reports are selected manually before publication;
- raw artifacts remain the numeric source of truth;
- p50, p95/p90, p99, sample count, filter, time window, and host provenance are shown;
- credentials and private endpoint identities are never published;
- gateway ACK is labeled as gateway health, not transaction landing;
- failed runs abort and remain visible to the operator rather than silently mixing retries into a result.