Supra achieves 500K TPS Throughput on 300 globally distributed nodes with sub-second Consensus Latency. Too good to be true? Nope, here’s how we do it: See thread below:

Supra achieves 500K TPS Throughput on 300 globally distributed nodes with Sub-Second consensus latency. Too good to be true? Nope, here’s how we do it: On Supra, we randomize nodes into subcommittees called Clans. We realize that data dissemination can be within a clan, form a data availability quorum certificate (DAQC) on a batch of 1000s of tx, and then only include the batch hash and the DAQC in our blockchain. Since Clans are randomized periodically and they are sized such that with high probability 51% of nodes in the Clan are honest while the entire network (aka Tribe) maintains 67% honest nodes — we can form a DAQC with only 51% of the nodes of the random subcommittee, while maintaining strong security and DA properties. Supra is designed to decouple data dissemination from consensus and also from execution. And each node can be a batch proposer (kind of like multi-leader designs others are trying to achieve) Our blocks remain relatively small because they only contain batch hash and daqcs. So our blocks are able to be distributed quickly throughout the network and our novel Moonshot Consensus finalizes blocks very quickly. Through this approach, Supra has achieved extreme speed throughput and sub-second consensus latency across 300 globally distributed nodes. Every node in the Tribe (active set of validator nodes) runs Moonshot consensus and requires a super-majority agreement. Now since Supra is designed to be MultiVM, each Clan is responsible for a specific VM or a set of VMs. Batches of txs are distributed to Clans on a VM basis accordingly. Clans will later propagate the data to a large set of RPC nodes so they can continue to be full nodes.

In summary, our major insight is: Data dissemination is expensive and a major constraint for individual nodes. We decouple data dissemination from Consensus and execution. There are more performance gains to be had by focusing on data dissemination vs. consensus vs. execution. All need to be optimized but the major opportunity for better blockchain performance is optimizing data dissemination through our unique tribe and clans network architecture. DAQC can be achieved with 51% of a randomized sub-committee and maintain strong security with high probability. Data doesn’t have to propagate to everyone immediately if the same sub-committee is responsible for executing the tx (useful especially since we provide MultiVM support) Each node can be a batch proposer this way, and our blockchain remains relatively tiny (only batch hashes and DAQC) So, blocks are lightweight and faster to disseminate to the entire network, thus achieving global ordering/consensus quicker. Hence we maintain great performance even beyond 300 nodes. DAGs also have each node propose blocks, but performance degrades extremely fast beyond 150 nodes and this is why, to my knowledge, we don’t see any DAG projects with large committees of node operators, thereby limiting their network’s decentralization.

Moonshot is simply beautiful ⚡️

@SUPRA_Labs 500K TPS is the cherry on top 🍒

@SUPRA_Labs LFMove bb !! ⛓️⚡️⛓️

@SUPRA_Labs It’s comin!!!!

@SUPRA_Labs

@SUPRA_Labs 500K TPS isn't to good to be true anymore, it's real! Supra team crack it let's go!

@SUPRA_Labs One of the most expected launches in 2024 !

@SUPRA_Labs That is indeed amazing! Good job 🫡 Sup sup 🔥