Skip to main content

One post tagged with "Sei"

Sei Network parallel-EVM Layer 1 blockchain

View all tags

Inside Sei V2's Parallel EVM: How 12,500 TPS Ships Today While Monad and MegaETH Race to Catch Up

· 10 min read
Dora Noda
Dora Noda
Software Engineer

In the parallel-EVM arms race that will define Layer 1 competition through 2026, one chain is shipping while the others are still benchmarking.

Sei Network's V2 mainnet has been quietly running optimistic parallel execution at a theoretical ceiling of 12,500 transactions per second with sub-400 millisecond finality since late 2024 — a full year before Monad's November 2025 mainnet launch and while MegaETH continues its specialized-node experiments. The question is no longer whether parallel-EVMs work. It's which architecture survives contact with the real workloads that come after the launch hype fades.

A 17,000-character technical teardown from Web3Caff Research traces Sei's path from a niche Cosmos SDK order-book chain in 2022 to the first production parallel-EVM L1, dissecting three interlocking innovations that make the throughput claims credible: optimistic parallel execution, Twin Turbo consensus, and SeiDB. But the same teardown also reveals the canonical gap every "high-TPS L1" eventually confronts — measured mainnet throughput sits at roughly 2,500-3,500 TPS under real dApp load, well below the 12,500 ceiling. Understanding what closes that gap, and what Sei's upcoming Giga upgrade does to push the ceiling toward 200,000 TPS, is the real story of where blockchain infrastructure is heading.

The Three-Pillar Architecture That Got Sei to Mainnet First​

Sei V2's performance does not come from a single breakthrough. It comes from three components engineered to compose, each attacking a different bottleneck in the legacy EVM stack.

Optimistic parallel execution is the headline feature, and it differs in a subtle but important way from Solana's Sealevel scheduler. Sealevel requires transactions to declare upfront which storage slots they intend to read or write, forcing developers to design around explicit dependency graphs. Sei's runtime takes the opposite approach: it speculatively executes all transactions in a block in parallel, tracks which state each transaction touches, and only re-executes the conflicting subset sequentially. Non-conflicting transactions clear in a single pass. The recursion continues until no unaccounted conflicts remain.

The trade-off is that optimistic execution wastes work when conflict rates spike — high-contention activity like a popular NFT mint or a single-pool DEX flash loan can degrade throughput as transactions stack up for re-execution. Monad uses a similar optimistic approach, while Aptos and Sui's Move-based parallel execution leans on resource-oriented programming to make conflicts statically analyzable. Each represents a different bet on how programmers will build at scale.

Twin Turbo consensus is what compresses Tendermint's notorious 6-second block times down to under 400 milliseconds. It's not a wholesale replacement of the underlying BFT engine — it's a suite of optimizations including aggressive timeout tuning, intra-block pipelining of proposal and voting phases, and a tight integration with the parallel-execution layer that lets transaction inclusion decouple from execution ordering. The result is single-slot finality at speeds previously associated with permissioned ledgers, while retaining the decentralization properties of a public BFT chain.

SeiDB is the least glamorous but arguably most consequential piece. The default Cosmos SDK uses an IAVL+ tree for state storage, which generates pathological disk I/O patterns under high write volume. SeiDB replaces this with a custom backend that splits state into two tiers — a write-optimized active layer and a read-optimized archive — reducing disk IOPS by roughly 10x according to Sei Labs' published benchmarks. When you're targeting tens of thousands of TPS, storage subsystem performance is no longer a footnote. It's the wall that breaks throughput before CPU does.

Geth Compatibility: The Strategic Choice That Mattered​

One architectural decision separates Sei V2 from Monad in a way that compounds over time: Sei imports Geth, the canonical Go implementation of the Ethereum Virtual Machine, directly into its node binary. Any Solidity smart contract deploys without modification. MetaMask, Hardhat, and Foundry work natively. Audit firms, tooling providers, and indexers built for Ethereum mainnet require zero adaptation.

Monad chose differently. Its team rebuilt the EVM from scratch in C++ to extract additional performance, accepting the long-tail cost of bytecode-level edge cases that may behave differently from canonical Ethereum. The bet pays off if Monad's performance advantage holds over time. It hurts if any of the thousands of audited Solidity contracts in production exhibit subtle execution differences when ported.

Sei's Geth-import strategy is what made the V2 launch survivable as a live network. It also made Sei the natural target for institutional deployments where compatibility risk is unacceptable — most visibly in January 2026, when Ondo Finance deployed USDY, the largest tokenized U.S. Treasury product by TVL, onto Sei mainnet. A tokenized Treasury issuer cannot tolerate edge-case EVM divergence. Geth-imports remove the question entirely.

The Mainnet Reality: 2,500 TPS, Not 12,500​

The empirical benchmarks tell a more complicated story than the marketing. Sei's mainnet currently sustains roughly 2,500 to 3,500 TPS under real dApp load — Astroport (the network's primary DEX), White Whale, Seiyans NFT activity, and the growing perpetual-futures market launched by Astroport Perps in December 2025. That figure sits well below the 12,500 TPS theoretical ceiling.

This gap is not a Sei-specific failure. It is the canonical gap every high-throughput L1 confronts when synthetic benchmarks meet production conditions. Three factors compress real throughput:

  • Conflict rates from real applications. Optimistic parallel execution rewards workloads with diverse state access patterns and punishes hot-state contention. A single dominant DEX pool routes most volume through a handful of pairs, and trades on the same pair conflict by definition.
  • Storage IOPS at saturation. Even with SeiDB's 10x improvement over IAVL, sustained write throughput above ~10,000 TPS pushes commodity NVMe drives into queue-depth territory where latency tail spikes degrade block times.
  • Validator network heterogeneity. Production validator sets span continents, latency varies, and Twin Turbo's tight timeouts assume favorable network conditions that don't always hold at the long tail.

Sei's TVL of roughly $560 million in DeFi (as of recent disclosures, with broader TVL exceeding $1 billion in June 2025) and 28 million active addresses tell the more important story: the chain is being used. The question is whether it can be used harder without breaking, which is exactly what the Giga upgrade aims to answer.

Giga: The 50x Bet That Defines Sei's 2026​

In December 2024, Sei Labs published the Giga whitepaper — a roadmap that, if delivered, would reset the entire L1 throughput conversation. Giga targets 5 gigagas per second of execution, which translates to approximately 200,000 to 250,000 TPS while preserving sub-400 millisecond finality. Devnet validation in 2025 hit 5.2 gigagas per second (~148,900 TPS) and 211 millisecond time-to-finality across a 20-validator set distributed across the U.S., Europe, and Asia Pacific.

Giga rebuilds three subsystems:

  • Autobahn consensus introduces multi-proposer block production, letting multiple validators propose disjoint transaction sets simultaneously rather than serializing through a single leader. This attacks the proposer bandwidth ceiling that limits single-leader BFT chains.
  • Asynchronous execution decouples transaction execution from block finalization entirely, letting the consensus layer commit ordering at one cadence while execution catches up at another. The pattern echoes what MegaETH attempts with specialized sequencer/prover/full-node roles.
  • A rebuilt EVM replaces the imported Geth with a performance-optimized implementation tuned for Sei's specific access patterns — closing the loop on the exact compatibility-vs-performance trade-off Sei avoided in V2.

The progressive mainnet rollout is scheduled throughout 2026, with the SIP-3 upgrade laying groundwork and full Giga deployment targeted by mid-year. If Sei pulls it off, the chain leapfrogs Monad's 10,000 TPS ceiling and approaches Web2-level transaction performance. If it doesn't, Sei's Geth-compatibility advantage gets eaten by Monad's mainnet maturity through the second half of 2026.

What This Means for the L1 Competitive Landscape​

The parallel-EVM category is no longer a research bet. It is an active competition with three live mainnets, distinct architectural choices, and visible institutional adoption. Sei has the production lead and the Giga roadmap. Monad has $269 million in fresh capital from its November 2025 ICO (85,820 participants, hosted by Coinbase) and a custom EVM built for raw speed. MegaETH ships node specialization that bets on a different scaling decomposition. Solana's Sealevel keeps grinding out 3,000-5,000 sustained TPS with $9B+ TVL but remains non-EVM.

The Move-based chains — Aptos and Sui — sit in a parallel category, betting that resource-oriented programming makes parallel execution strictly better than any retrofit onto Solidity semantics. They've shipped to mainnet and have working ecosystems, but the gravitational pull of EVM tooling makes the parallel-EVM lane the more contested one.

What the Sei deep dive ultimately reveals is the architectural ceiling every parallel-execution chain will eventually hit: above approximately 10,000 sustained TPS, storage IOPS becomes the binding constraint, not VM parallelism. This is why Giga puts as much weight on the storage layer redesign as on consensus. It's also why the next frontier of L1 scaling — already visible in early 2026 conversations — is shifting from "parallelize the VM harder" to state-sharding combined with data-availability composition. Sei is positioned to lead that transition because it has already shipped one parallel-EVM and is iterating on the second.

The Infrastructure Layer Underneath​

For developers building on Sei, Monad, or any parallel-EVM in 2026, the infrastructure question gets more nuanced than it was on legacy Ethereum. Optimistic execution means transaction ordering depends on conflict resolution, which means RPC providers need to expose the right primitives for builders, sequencers, and indexers to make sense of execution traces. Sub-400ms finality is meaningless if your indexer is 30 seconds behind, and 12,500 TPS amplifies any reliability gap in the read path.

The chains that win the parallel-EVM era will be the ones whose infrastructure ecosystem keeps up — RPC reliability, archive node coverage, indexer freshness, and the kind of multi-chain abstraction layer that lets a developer treat Sei, Monad, and Solana as substitutable rather than separate integrations.

BlockEden.xyz provides enterprise-grade RPC and indexing infrastructure across Sei, Solana, Sui, Aptos, Ethereum, and the broader L1 landscape. As parallel-EVMs mature from testnet promises to production workloads, explore our API marketplace to build on infrastructure designed for the throughput frontier.

The Bottom Line​

Sei V2 is the proof point that parallel-EVMs can ship to mainnet, support real institutional deployments like Ondo's USDY, and run live workloads at 2,500-3,500 sustained TPS — not the 12,500 TPS marketing number, but a production figure that already exceeds Solana's sustained throughput while running unmodified Solidity contracts. Whether Sei holds that lead depends on Giga delivering its 5 gigagas-per-second target before Monad matures and MegaETH proves its specialized-node thesis.

The 2026 throughput race is no longer about benchmarks. It's about which architecture composes cleanly with the storage, consensus, and DA primitives that define the next phase of L1 design. Sei got there first. The next twelve months decide whether first-mover advantage in parallel execution converts into durable category leadership.

Sources​