Key Highlights:
- Celestia’s Fibre network sustained 3.07 Tb/s of confirmed onchain blob throughput in a full-pipeline benchmark.
- The best 60-second window averaged 3.69 Tb/s, while the best 30-second window reached 4.27 Tb/s.
- Celestia says the test demonstrates Fibre’s ceiling potential, while mainnet will initially use smaller blobs and different capacity settings.
Celestia Fibre Reaches 3.07 Tb/s in Full-Pipeline Test
Celestia’s Fibre data-availability network sustained 3.07 terabits per second (Tb/s) in a benchmark that tested the complete operating pipeline rather than a limited laboratory simulation. The test covered encoding new blobs, distributing and storing the resulting pieces, collecting validator signatures, and committing the data onchain.
The result exceeds the 1 Tb/s performance Celestia previously demonstrated with Fibre in a partial test earlier this year. Celestia said a prospective customer with unusually demanding throughput requirements prompted the company to measure how much capacity Fibre could sustain under real operating conditions.
Fibre performed above its overall average during its strongest periods. The best 60-second interval averaged 3.69 Tb/s, while the best 30-second interval reached 4.27 Tb/s. Celestia said the reported figures include only blob data confirmed onchain. They exclude recovery pieces created during encoding and the traffic required to distribute pieces to validators, meaning total network activity would be higher.
How Fibre Separates Blob Data From Blockchain Commitments
Fibre’s performance is based on separating the movement of blob data from the work performed by Celestia’s blockchain. Blob data is sent directly to validators, while the chain records a commitment to that data and the validators’ signatures. The blockchain therefore does not process the full underlying data volume; it processes proof that the data exists and has been verified.
Under the system, a client encodes a blob and sends each validator its assigned share of pieces. Validators verify and store those pieces before signing to confirm receipt. After validators representing two-thirds of voting power have signed, the client submits the signatures to Celestia together with the blob commitment.
Celestia’s engineering team spent months improving the pipeline. Since March, the company has reduced duplicate verification, reused memory, and removed unnecessary data copies so Fibre could process more uploads simultaneously.
ARM Vectorization Cuts Blob Encoding Time
One of the largest performance improvements came from Reed-Solomon encoding. Before a blob is sent, Fibre divides it into pieces and creates recovery pieces so the original data can be reconstructed if some pieces are lost during transmission. Because every blob goes through this process, even small inefficiencies can significantly affect total throughput.
The ARM chips used in testing, including AWS Graviton machines, can process multiple values in a single instruction. However, Fibre’s original Reed-Solomon implementation processed data one value at a time. Celestia engineers used NEON, ARM’s vector instruction set, to create vectorized kernels and optimize the workload for the hardware.
That work reduced the median time required to encode a 2 GiB blob from 6.8 seconds to 0.9 seconds, an improvement of approximately 7.5 times.
Chain Processing and Storage Changes Remove Bottlenecks
Once encoding and distribution became faster, Celestia’s chain-processing capacity became the main constraint. In testing, uploads required approximately one second for encoding and 0.5 seconds for piece distribution, but confirmation could take as long as 14 seconds because data was arriving faster than the chain could process it.
Celestia addressed the bottleneck by caching and reusing successful signature checks instead of repeating them. The company also parallelized verification and streamlined validator processing for PayForFibre (PFF) transactions, the onchain messages that record Fibre uploads. These changes reduced the time needed to validate a block proposal from scratch, without a cache, from 10.35 seconds to 1.12 seconds.
The company then increased the per-block capacity for Fibre submissions from 200 to 2,000 and shortened block time to one second. Those changes reduced the average confirmation wait to 2.2 seconds in the final benchmark.
Storage infrastructure required a separate adjustment. Celestia initially used network-optimized AWS instances because of their bandwidth, but the attached Amazon EBS disks could not write data as quickly as it arrived. The company moved blob-piece storage to Amazon S3, AWS’s object-storage service.
S3 performance began to decline at roughly 3,500 uploads per second. To reduce the number of requests, Celestia grouped 16 Fibre pieces into each object and distributed writes across hash-based key groups and multiple buckets. Together with the encoding and chain-processing improvements, these storage changes enabled the 3.07 Tb/s benchmark.
Fibre Mainnet Performance Will Differ From the Benchmark
Celestia cautioned that several benchmark conditions will not apply directly when Fibre launches on mainnet. The test used 2 GiB blobs, while mainnet will initially impose a 128 MiB limit. The benchmark also used one-second blocks and a maximum of 2,000 Fibre submissions per block, compared with the current limit of 200.
Memory and concurrency settings were calibrated for the specific AWS machines used in the test. The run also relied on an experimental performance branch of Fibre’s production pipeline, meaning some of the reported optimizations are still being refined before a broader rollout.
As a result, the benchmark is best understood as a demonstration of Fibre’s potential under tuned conditions rather than a guarantee of its out-of-the-box mainnet performance. Celestia plans to match initial mainnet capacity to early usage and expand it as application demand develops.
Why This Matters for Blockchain Data Capacity
Celestia argues that demand for blockchain blockspace could soon exceed the capacity available from existing systems. The company points to AI agents making payments on behalf of users and the broader movement of global financial activity onto blockchains as potential sources of growing demand.
According to Celestia, Fibre operating at 3 Tb/s could support an AI agent for every person on the planet, with each agent submitting one transaction every four seconds. The company’s stated objective is to scale Fibre’s blockspace capacity to 3 Tb/s and beyond as application demand increases.
The benchmark was led by Vlad Krinitsyn, with contributions from Rachid Chami, Preston Evans, Hlib Kanunnikov, Alex Kiss, Rene Lubov, and Rootul Patel.
Frequently Asked Questions
What throughput did Celestia Fibre achieve?
Fibre sustained 3.07 Tb/s of blob data confirmed onchain during the full-pipeline benchmark. Its strongest 60-second window averaged 3.69 Tb/s, and its strongest 30-second window reached 4.27 Tb/s.
What did the Fibre benchmark test?
The test covered blob encoding, piece distribution and storage, validator signature collection, and final onchain commitment. It was designed to measure the complete pipeline rather than a partial or laboratory-only operation.
Will Fibre launch on mainnet at 3.07 Tb/s?
Not necessarily. Celestia said the benchmark used 2 GiB blobs, one-second blocks, and a limit of 2,000 submissions per block, while initial mainnet settings will include 128 MiB blobs and different capacity parameters. The result represents tuned ceiling potential rather than a guaranteed day-one mainnet throughput.