The evolution of blockchain technology from the early stages of Bitcoin to the explosion of decentralized applications (dApps) on Ethereum has posed a difficult problem of scalability. For years, the industry has focused on optimizing consensus algorithms to achieve higher transaction speeds. However, according to analysis from Tan Phat Digital, when entering the period 2024-2026, the research focus has shifted strongly from the consensus layer to the implementation layer. The core question is: Does parallel execution really solve the speed problem? The short answer is yes, but the technical reality is much more complex. Applying parallel execution is like adding workers to a construction site; it is only truly effective if the work can be broken down and the overhead does not cancel out the productivity benefits.
Theoretical framework for parallel execution and concurrency in distributed systems
To evaluate the effectiveness of parallel execution, it is necessary to clarify the difference between two often confused concepts: concurrency and parallelism. Concurrency refers to the ability to process multiple tasks by overlapping execution times, much like a chef making multiple dishes by slicing meat while waiting for water to boil. In contrast, true parallelism requires the simultaneous execution of tasks on separate physical resources, similar to having two chefs cooking two different dishes in two separate kitchens.
In traditional blockchain architectures like the early Ethereum Virtual Machine (EVM), every transaction is processed sequentially. This ensures determinism but creates a single queue, where one complex transaction can clog the entire network. Parallel execution breaks this model by allowing unrelated transactions to be validated at the same time across multiple CPU or GPU cores.
Optimal conditions for parallel execution to work
Parallel execution is not a "silver bullet" for every type of task. Its performance depends on the nature of the workload being processed. Empirical studies show three main scenarios where parallelization provides outstanding benefits:
State-independent tasks: Jobs where the result of one part does not affect the other, such as processing payment transactions between completely different account pairs.
Large-scale data processing (Big Data): When the system needs to process millions of records writes, distributing the workload across thousands of threads significantly reduces overall completion time.
Compute-Intensive: Scientific problems, AI training, or deep learning require massive parallel computing capabilities on specialized hardware such as NVIDIA GPUs.
Characteristics comparing sequential and parallel processing parallel
Sequential processing:
Execution structure: One task at a time in a fixed order.
Determinism: Very high and easy to predict results.
Hardware utilization: Low, usually only uses a single CPU core.
Latency: Increases proportionally to number of transactions in the queue.
Scope of application: Algorithms have a strong dependence on previous data.
Parallel processing:
Execution structure: Many tasks take place simultaneously on many processing cores.
Determinism: Complex, needs state locking mechanisms or post-execution authentication.
Hardware utilization: High, taking full advantage of multi-core architecture and modern GPUs.
Latency: Significantly reduced for independent tasks.
Application scope: Suitable for Big Data, AI and independent blockchain transactions.
See more: What is Blockchain Trilemma?
Mathematical and technical barriers: The "black hole" of speed
Although parallel execution promises processing speeds of tens of thousands of transactions per second (TPS), in reality implementation always faces strict theoretical limits, especially Amdahl's Law.
Amdahl's Law and the limits of acceleration
Amdahl's Law states that the maximum speed of a system is limited by the portion of the program that must run sequentially. If a block of source code has 10% of the parts that cannot be run in parallel, then even if the system has an infinite number of CPUs, the speed will never be more than 10 times faster than the original.
Effect of parallelization ratio on maximum speed
50% parallel ratio: Maximum speed increase by 2 times.
Parallelization ratio 75%: Maximum speed increase 4 times.
90% parallel ratio: Maximum speed increase 10 times.
95% parallel ratio: Maximum speed increase 20 times.
99% parallel ratio: Maximum speed increase 100 times times.
Overhead and Resource Contention
Dividing tasks for parallel execution is not free. The system must spend resources to distribute work to threads, monitor status, and aggregate results. If the tasks are too small, the time spent on this management can take longer than the sequential task execution time. Resource contention occurs when multiple threads compete for access to the same memory area, creating bottlenecks that reduce overall system efficiency.
Parallel execution models in modern Blockchain architecture
Tan Phat Digital has synthesized three main models that are reshaping industry performance:
1. Deterministic Model (Solana Sealevel)
Solana requires each transaction to declare in advance all the accounts it will read or write to. Thanks to this information, the system schedules simultaneous execution of transactions that do not overlap accounts. The advantage is extremely high performance but the burden is on the developer to manage complex state access lists.
2. Optimistic Collision Control Model (Aptos & Sei)
Instead of declaring it upfront, the system assumes most transactions are conflict-free and run in parallel immediately. Then the new system checks for conflicts; If so, the transaction will be rolled back and re-executed. Sei Giga, a 2026 upgrade, reached 200,000 TPS in a test environment thanks to this mechanism.
3. Object Oriented Model (Sui)
Sui changes the data structure to an object model. Transactions that interact with private assets (Owned Objects) can bypass global consensus, allowing for instant execution with extremely low latency. The new Stingray system on Sui also allows reaching 20,000 TPS for highly commutative tasks.
In-depth analysis: Monad and the EVM parallelization revolution
By 2026, Monad becomes the symbol for optimizing the execution layer while remaining 100% compatible with Ethereum. Monad's architecture solves the biggest bottleneck: accessing data from disk.
MonadDb: State-aware storage system, supporting asynchronous I/O, reducing state access latency from 10ms to less than 1ms.
Asynchronous execution: Separation of consensus and execution, allowing the CPU to never be in a state idle while waiting for the network to transmit data.
Berachain BeaconKit: Another technology framework is also helping to reduce block generation time by 40% through parallel execution and optimistic payload construction.
Comparing 2026 forecast performance metrics
Sei Giga: Actual TPS hits 12,500+ (with a record of 200,000+ on devnet). Block creation time is 0.4s and completion latency is less than 0.4s. Using Multi-proposer mechanism.
Monad: Actual TPS reaches 10,000. Block creation time 1.0s and completion delay ~1.0s. Uses Optimistic mechanism combined with MonadDb.
Sui: Actual TPS from 2,000 - 4,500. Block creation time is 0.4s and completion latency is less than 0.4s. Uses Object Oriented mechanism (Mysticeti).
Solana: Actual TPS from 2,500 - 5,000. Block creation time 0.4s and completion delay ~2.0s. Using Deterministic mechanism (Sealevel).
Aptos: Actual TPS from 1,500 - 3,500. Block creation time is 0.7s and completion delay is ~0.9s. Using the Optimistic mechanism (Block-STM).
Ethereum's development roadmap: From sequential to parallel
Ethereum is actively implementing important upgrades through the Glusterdam and Hegota roadmap in 2026.
- more.
Verkle Tree: Reduces the size of state proofs, allowing customers to more efficiently parallelize block validation.
10 Detailed Case Studies on Parallel Execution (Updated 2026)
Below is a detailed analysis from Tan Phat Digital on typical projects leading the charge performance racing:
1. Solana (Sealevel)
Solana is a pioneering project in using deterministic parallel execution through the Sealevel engine.
Mechanism: Requires transactions to declare in advance the accounts they will access (Read-Write set).
Performance: Reach 2,500 - 4,000 actual TPS on mainnet in 2024-2025, with the ability to handle up to tens of thousands of TPS under ideal conditions.
Advantages: Take full advantage of multi-core GPUs and CPUs.
2. Monad (MonadDb & Asynchronous Execution)
Monad represents a quantum leap for the EVM ecosystem in 2026.
Mechanism: Combines Optimistic Execution with the MonadDb custom database to eliminate the I/O bottleneck.
Performance: Reaches 10,000 TPS consistently with 0.4 second block generation time.
Highlights: 100% bytecode compatible with Ethereum, allowing dApps to move without code modifications.
3. Sei Giga (Multi-Proposer Consensus)
Sei Network's Giga upgrade is one of the most important milestones of 2026.
Mechanism: Allows multiple validators to propose blocks simultaneously in a consensus round, combined with automatic parallel execution.
Performance: Reaches a record of 200,000 TPS on devnet and maintained over 12,500 TPS on actual mainnet.
Latency: Finality reaches extremely low levels, under 400ms.
4. Sui (Object-Centric Fast Path)
Sui uses a data model that is based on objects instead of accounts.
Mechanism: Divide transactions into Fast Path (for Owned Objects) and Consensus Path (for shared objects).
Performance: Ability to process more than 65 million transactions in a peak day.
Advantages: Completely eliminates the consensus step for most simple transactions, providing a sense of instant feedback.
5. Aptos (Block-STM)
Aptos uses a sophisticated Software Transactional Memory mechanism.
Mechanism: Block-STM automatically detects and resolves conflicts during execution, rolling back only affected transactions instead of the entire block.
Performance: Actual TPS remains at level 1,500 - 3,000 TPS.
Advantages: Developer-friendly because the system automatically handles parallelism.
6. Berachain (BeaconKit)
Berachain delivers "EVM-Identical" architecture with outstanding performance through BeaconKit.
Mechanism: Separation of execution layer and consensus layer, allowing optimistic payload construction (parallel block construction with voting process).
Performance: 40% reduction in time block creation compared to traditional EVM chains.
Highlights: Proof-of-Liquidity (PoL) mechanism helps align network security with ecosystem liquidity.
7. Hyperliquid (HyperBFT)
Hyperliquid is a good example of blockchain optimization for specific financial applications (App-chain).
Mechanism: Uses a custom HyperBFT consensus algorithm and a completely on-chain order book (CLOB).
Performance: Ability to process 200,000 orders per seconds with a delay of less than 1 second.
Application: Specialized for derivatives and perpetual trading with zero gas costs for trading users.
8. BNB Chain (EIP-7928 & Reth)
BNB Chain implements a "Smarter Execution" strategy to maintain its position as the most popular transaction chain.
Mechanism: Apply EIP-7928 (Block-level access lists) and migrate to Rust-based Reth client.
Performance: Target achieved above 20,000 TPS by 2026 under real-world conditions.
Optimum: Reduce gas fees to extremely low levels (about 0.05 gwei) thanks to improved execution efficiency.
9. Polygon (AggLayer & Gigagas)
Polygon transforms itself from a single Sidechain into a network of perfectly interoperable chains through AggLayer.
Mechanism: Focuses on "Gigagas" throughput (billions of gas units per second) through fully parallelized execution on Polygon PoS.
Effectiveness capacity:Roadmap to 100,000 TPS through combining ZK-proofs and parallelization.
10. Altius (Execution-First Architecture)
Altius is an emerging project focused on handling real-world tasks rather than theoretical benchmarks.
Mechanism: Redesign the entire execution flow to prioritize determinism and high contention load tolerance.
Strengths: Benchmark is performed on real-life scenarios such as NFT Minting or DEX Swapping instead of simple money transfer transactions.
Frequently Asked Questions (FAQs) about parallel execution
What is Parallel Execution? Is the ability to process multiple transactions simultaneously or tasks on multiple CPU/GPU cores instead of processing them one by one sequentially, helping to optimize hardware resources and increase network speed.
Why does a traditional blockchain like Ethereum run sequentially? To ensure determinism (every network node produces identical results). Sequential processing helps avoid data conflicts but creates a "bottleneck" when the transaction volume is large.
How does Amdahl's Law affect blockchain speed? It states that the maximum speed is limited by the portion of transactions that are required to run sequentially (for example, transactions interacting with the same hot account). If 10% of the workload is sequential, the speed cannot increase more than 10 times regardless of how many CPUs are present.
What is State Conflict? Occurs when two or more transactions want to change data of a single account at the same time. The system must rearrange them to avoid data errors.
What is the difference between Solana and Aptos in parallel processing? Solana requires declaring the account list in advance (Deterministic), while Aptos allows running first and checking for conflicts later (Optimistic - OCC).
What is Sei Giga and how fast is it? Sei Giga is the 2026 upgrade of Sei Network, reaching 200,000 TPS on the devnet thanks to its multi-proposer architecture and automatic parallelization mechanism.
What problem does MonadDb solve? It solves the I/O (drive read/write) bottleneck. MonadDb reduces state access time from 10ms to less than 1ms by prefetching data and supporting asynchronous I/O.
What are the advantages of "Owned Objects" in Sui?Owned Objects belong to only one address, allowing related transactions to skip the complicated consensus step and execute almost instantly (Fast Path).
What role does EIP-7928 play in the Ethereum 2026 roadmap? It provides a "Block Level Access List" (BAL), which records all accounts and storage locations accessed, helping network nodes perform data reading and transaction validation in parallel.
Does parallel execution reduce gas fees? Yes, because it increases processing capacity. network management, reducing queue congestion, thereby reducing the pressure to increase fees to receive priority processing.
Why does the hot NFT mint slow down parallel chains? Because thousands of users interact with a single NFT contract (hot account), the system has to return to sequential processing to ensure the consistency of the total supply counter.
What are the hardware requirements for validators of these chains? Usually very high requirements: 64 core CPU or more, minimum 128GB RAM and high-speed NVMe hard drive to handle huge data streams.
Is Monad fully compatible with Ethereum? Yes, Monad maintains 100% compatibility at the bytecode level, allowing developers to deploy existing Solidity contracts without modifying the source code.
What is the difference between "Concurrency" and "Parallelism"? Concurrency is managing many things at the same time (possibly overlapping time), while Parallelism is actually doing many things at the same time on separate hardware resources.
When does Ethereum's "Glamsterdam" Roadmap launch?Slated for activation in the first half of 2026 (around May or June), focusing on layer 1 execution efficiency and improving the MEV mechanism.
Parallel execution is no longer an option but has become a mandatory architectural requirement for any network that wants to reach global adoption. Tan Phat Digital believes that it truly solves the speed problem by breaking the traditional sequential processing barrier.
However, its effectiveness is not limitless. Amdahl's Law is always a reminder of the limits of sequential logic. The success of projects like Monad, Sei or Sui shows that the future lies in end-to-end optimization: from the state storage layer, consensus mechanism to redesigning the data model to minimize conflicts. For developers, parallel execution provides a wider "highway", but to move quickly, the application still needs to be intelligently designed to avoid data "collision" on available lanes.
Share
