The Hyperliquid L1 will be transitioning to HyperBFT, a custom consensus algorithm written from the ground up. The implementation is in final stages of testing on testnet. The mainnet migration is expected to happen in the coming weeks.
The Hyperliquid L1 is one of a few chains in production with real demand above 20k TPS. With a scope expanding far beyond perps trading (more on that later), it is crucial to optimize the foundational consensus algorithm.
Like most modern consensus algorithms, HyperBFT is based on Hotstuff and subsequent improvements. Huge thanks to the researchers who continue to push the boundaries of what is possible on a decentralized state machine.
Some properites of HyperBFT:
1) TPS will be massively improved. The current bottleneck for the Hyperliquid L1 is Tendermint, which caps out at 20k orders/second. HyperBFT can support up to 100x the throughput, though in practice the state machine execution will be a bottleneck at around 200k orders/second.
2) Consensus does not block on execution. Unlike Ethereum and Tendermint, the consensus algorithm can continue sequencing transactions without waiting for the hash of the current block being executed.
3) Confirmation latency will improve and stabilize, with block times bounded only by the network delay. Unlike earlier BFT algorithms such as Tendermint, there is no synchronous timescale baked into the consensus algorithm itself under normal block production. Blocks are produced as quickly as a quorum of validators can communicate. This property is sometimes called "optimistic responsiveness."
The mainnet rollout will be incremental, focusing first on stability at performance parity with the current L1 before cranking up throughput.