← Workload Catalog

Distributed State

Replica state convergence

Replicas converge to identical state by exchanging deltas — no coordinator, no consensus, a fraction of the bandwidth.

CONVERGEScoordination-free convergence + ~92% less data moved
Convergence (correctness)PASS
replicas reach identical state with no coordinator — byte-identical across 1/2/4/8 banks (0x17e9fe29b5dc0aad)
any node, any order, after any partition → identical merged state. The same property, on silicon.
Data moved (bandwidth)PASS
2048 B/tick full-state → 160 B/tick delta (measured) = ~92% less
at ~5% of state changed per tick — scales with change density, not a universal multiplier

Measured on the AX7020 ATOMiK adapter: an edge-telemetry sync tick moved 160 B vs 2,048 B of full state, with exact 64-bit LOAD/ACCUM/READ round-trips on a falsification-tested harness. Convergence is the same order-independence proven on the AX7020 (2026-06-21); reproduce with abench 65537 0xDEADBEEF0BADF00D.

Run this on our silicon → sign conditional on your bar

This is you if

  • You replicate state across edge / multi-region nodes or to client caches.
  • You use (or are weighing) CRDTs or consensus to keep replicas consistent.
  • Sync bandwidth and coordination latency are real costs for you.

The workload

  • State is replicated across N nodes (edge, multi-region, or client-side).
  • Nodes update concurrently and must converge to the same state.
  • Updates and reconciliations cross the network — bandwidth and latency matter.
  • No node is authoritative; there is no global order.

Today — coordinated / full-state sync

Keeping replicas consistent means a coordinator / consensus (Raft / Paxos: latency, a leader bottleneck) or CRDTs (metadata overhead, and you still ship state or fat op-logs). And each sync tends to ship full state or a growing op-log, not just what changed.

With ATOMiK — order-independent, coordination-free

Replicas exchange delta-state. Because the merge is order-independent and self-consistent — byte-identical no matter which deltas arrive in which order at which node — replicas converge with NO coordinator and NO consensus, and each tick ships only the delta, not the full state.

  • Coordination-free convergence — no leader, no consensus round-trip.
  • Any node, any order, after any partition → identical state (the same byte-identical property proven on silicon).
  • Ships the delta, not full state — measured 160 B vs 2,048 B (~92% less) at ~5% change density.
  • Reconnect after a partition = replay deltas in any order and converge — no reconciliation protocol.
Want to test a different rate or operator? Open the interactive benchmark tool · Representative archetype, not a named customer. Numbers are measured engine facts; the value is lock-elimination + the byte-identical guarantee.