Introduction
Cerberus is Radix's consensus protocol, published as a peer-reviewed paper in 2020 and reviewed in the Journal of Systems Research (2023). Independent reviewers concluded that Cerberus is among the most efficient multi-shard consensus protocols in terms of throughput and latency. This article compares it against established BFT protocols.
Single-Pipeline Protocols
PBFT (Practical Byzantine Fault Tolerance)
PBFT (Castro & Liskov, 1999) is the foundational BFT protocol. It runs a three-phase commit process (pre-prepare, prepare, commit) requiring O(n²) message complexity. PBFT provides deterministic finality but is limited to small validator sets (~20-50) due to communication overhead. All transactions are processed sequentially through a single leader.
HotStuff
HotStuff (2018) reduces PBFT's message complexity to O(n) per round through a pipelining approach with rotating leaders. It forms the basis for Meta's abandoned Diem/Libra and the Aptos blockchain. HotStuff is more scalable than PBFT for larger validator sets but remains a single-pipeline protocol ā all transactions are totally ordered.
CometBFT (Tendermint)
CometBFT (formerly Tendermint) is a BFT protocol used across the Cosmos ecosystem. It provides instant finality (no probabilistic confirmation) with O(n²) communication complexity. Each Cosmos chain runs its own CometBFT instance, enabling the multi-chain architecture but limiting individual chain throughput.
DAG-Based Protocols
Narwhal/Bullshark
Narwhal (2021) separates data dissemination from consensus ordering using a DAG (Directed Acyclic Graph) structure. Bullshark (2022) adds partially synchronous BFT ordering on top. This combination achieves high throughput by parallelising data availability but still requires total ordering of all transactions ā a single-shard model. Narwhal/Bullshark formed the basis of Sui's consensus layer.
Cerberus: Parallelised BFT
Cerberus takes a fundamentally different approach. Rather than optimising a single consensus pipeline, it runs a proven BFT process in parallel across practically unlimited shards, each maintaining its own ordered transaction stream. Cross-shard transactions are handled through "braiding" ā a synchronisation mechanism that achieves atomic commitment across involved shards without global coordination.
| Property | PBFT | HotStuff | CometBFT | Narwhal/Bullshark | Cerberus |
|---|---|---|---|---|---|
| Shard Support | Single | Single | Single (per chain) | Single | Unlimited |
| Cross-Shard Atomicity | N/A | N/A | Via IBC (async) | N/A | Native (braided) |
| Message Complexity | O(n²) | O(n) | O(n²) | O(n) | O(n) per shard |
| Finality | Deterministic | Deterministic | Deterministic | Deterministic | Deterministic |
| Throughput Scaling | Fixed ceiling | Fixed ceiling | Fixed per chain | Fixed ceiling | Linear with nodes |
The key innovation is that Cerberus provides atomic composability across shards ā something no other production BFT protocol achieves ā while maintaining linear throughput scaling.
External Links
- Cerberus: A Parallelized BFT Consensus Protocol ā arXiv
- What is Cerberus? ā Radix Knowledge Base
- PBFT ā Wikipedia
- HotStuff: BFT Consensus in the Lens of Blockchain ā arXiv
