Scaling Research Watch / Network & Protocol
In Brief
A revised Teranode scaling study posted to SSRN on August 10 reports a measured 79.09 billion pipeline-processed transactions per second across 100 processing instances spanning 10 geographic regions. The experiment used controlled conditions designed to measure the limits of Teranode’s horizontally scalable transaction pipeline, while separate tests reported zero accepted double spends across 520 million UTXOs. The paper explicitly distinguishes these measurements from open-network BSV Blockchain throughput and from lower production-oriented estimates that remain to be validated experimentally.
News Report
A revised research paper examining the horizontal scaling of Teranode reports a measured aggregate throughput of 79.09 billion pipeline-processed transactions per second across a 100-instance test environment.
The 47-page paper, Horizontal Scaling of UTXO-Based Transaction Processing: Architecture, Empirical Validation, and Fleet-Scale Projection, was posted as a new version on SSRN on August 10. An earlier version had been posted in April.
The study evaluates Teranode as a horizontally scaled transaction-processing architecture in which functions traditionally handled inside a monolithic blockchain node are separated into 14 microservices. Those services cover transaction ingestion, validation, UTXO processing, block assembly and validation, persistence, peer-to-peer communication and other node functions.
The reported 100-instance experiment spans 10 geographic regions. At the measured high-end configuration, the study reports aggregate throughput of 79.09 billion TPS, with a scaling efficiency of 0.783 relative to the underlying per-instance measurements.
The meaning of “TPS” is narrower than complete public-network throughput and is carefully defined in the paper. A transaction is counted after moving through the experimental processing pipeline, including ingestion, script validation, UTXO spend processing, Kafka offset commitment, Merkle-tree assembly, block-header commitment and header propagation within the Teranode fleet.
The experiment did not measure transaction propagation between independent miners across the open internet. The processing instances operated inside a single administrative trust domain, and the study excluded Byzantine behavior such as deliberately malicious validators or message fabrication.
The headline experiment also used deliberately favorable processing conditions. Transactions were synthetic P2PKH payments averaging about 500 bytes, the experimental UTXO set was small enough to remain in DRAM, transaction injection used a compact gRPC representation, and proof-of-work difficulty was reduced so mining would not become the bottleneck being measured.
These qualifications are important because the paper is testing whether the underlying transaction-processing architecture can scale horizontally, rather than claiming that BSV Blockchain mainnet is presently processing 79.09 billion transactions per second.
The study provides several additional figures that help place the headline result in context.
A more conservative measured distributed baseline corresponds to approximately 61.2 billion TPS across the 100-instance fleet. A separate 24-hour single-instance endurance test recorded a minimum throughput of 841 million TPS, which the paper translates into a 65.8 billion TPS fleet-scale sustained floor for capacity-planning purposes.
The authors then model how performance could change when additional production conditions are introduced. Accounting independently for larger UTXO storage spilling from memory to NVMe, full peer-to-peer ingress overhead and a mixture of more complex transaction types produces an analytical estimate of approximately 36–42 billion TPS at 100 instances.
The paper subsequently applies an interaction correction because those performance penalties may influence one another rather than acting independently. That lowers the central production-oriented estimate to approximately 33 billion TPS, with a wide 95% interval of roughly 19–48 billion TPS.
Those lower figures are estimates rather than additional measured benchmarks. The paper identifies empirical testing with production-scale NVMe storage, mixed transaction workloads and full open-network peer-to-peer ingress as important remaining work.
The study also tested UTXO consistency under deliberate conflicting transactions. Across the experimental campaign, it reports 520 million UTXOs observed and 12.8 million deliberate double-spend attempts detected and rejected, with zero double spends accepted.
Teranode’s tested architecture uses Kafka to order transaction-processing work and a shared Aerospike UTXO store using generation-checked compare-and-swap operations. When two transactions attempt to spend the same output, the state change associated with one transaction prevents the conflicting operation from succeeding under the experiment’s crash-fault model.
Fault handling was tested separately by deliberately throttling a Kafka broker during sustained processing. At the 100-instance scale, throughput fell by approximately 43% during the disruption. After the fault cleared, the system automatically settled at approximately 82–91% of its previous steady-state throughput, with recovery to that post-fault state taking around 100 seconds.
The UTXO safety audit continued to report zero accepted double spends during the tested fault conditions. The paper nevertheless limits that result to its stated trust model: components may fail and restart, but deliberately malicious behavior by processing participants was outside the scope of the experiment.
Another finding is that extreme throughput begins to move the scaling problem away from computation and toward data retention. The paper calculates that retaining the complete history generated at the 79.09 billion TPS rate would produce approximately 1.25 zettabytes of data per year.
Its proposed operating model therefore relies heavily on pruning spent transaction data while retaining the active UTXO set, recent data required for reorganizations and the block-header chain. Under the study’s assumptions, pruning reduces the corresponding storage requirement at the 100-instance scale to approximately 4.75 exabytes.
The paper also discusses scaling beyond 100 instances, including a proposed transition to sharded UTXO storage. Those larger configurations have not been experimentally validated. The study explicitly categorizes fleet sizes beyond the tested 1–100-instance range and trillion-TPS figures as projections or design-envelope estimates rather than demonstrated Teranode performance.
The August revision therefore provides several different kinds of evidence that should not be collapsed into a single number: measured transaction-processing performance up to 100 instances, measured endurance and fault behavior, analytical estimates for more demanding operating conditions, and projections for architectures that remain future work.
BSV TIMES Read
The 79.09 billion figure is striking, but the more important result may be what the study is beginning to show about the architecture underneath it. Teranode is being tested not simply by making one machine faster, but by distributing transaction processing across many parallel instances while preserving a consistent UTXO state. The paper also makes unusually clear where measurement ends and projection begins. Open-network ingress, larger production UTXO sets, mixed transaction workloads and architectures beyond 100 instances remain important tests ahead. If those stages continue to validate the horizontal-scaling model, the significance will be less about one record number than about demonstrating that BSV Blockchain capacity can expand by adding processing infrastructure rather than repeatedly redesigning the base protocol.
Source Links
BSV Blockchain — Teranode GitHub Repository
Posted on August 17, 2026
Leave a comment