High-performance S3-compatible object storage
Over 1.3 Million Object/s with ~3 ms P99 latency and native atomic rename. Deploys to your own AWS or GCP account in under 5 minutes.
$ just deploy create-vpc \
--template perf_demo --with-bench
✓ cluster deployed in under 5 minutes
# benchmark — 4 KiB objects, single bucket, 1 meta node
GET 1.3M Obj/s P99 3.4 ms
PUT 280K Obj/s P99 8.5 msWhy FractalBits
Built for Performance
FractalBits pairs a modern storage architecture with low-level systems engineering to outperform AWS S3 Express One Zone — at a much lower cost.
1.3M+ Object/s
Over 1.3M 4 KiB reads/s at ~3ms P99 — built for AI training and data analytics.
Fractal ART Metadata Engine
Full-path design skips the distributed transactions of inode-based systems — scalable, with full directory semantics and atomic rename.
Multi-Protocol Access
One backend, many interfaces — the same data over S3 object API and POSIX file system, with more protocols coming.
NVMe-Native Data Engine
SSD-optimized data engine with a low-overhead I/O path that scales directly with your NVMe hardware — no EBS required.
Linear Scale-Out
Outgrow a single node? Add metadata and data nodes to scale throughput and capacity near-linearly — no re-architecting.
5-Minute BYOC
Deploy to any AWS or GCP region with a single command — running in under 5 minutes.
Powered by
Use Cases
Built for Demanding Workloads
Wherever small objects, low latency, and atomic operations decide whether infrastructure keeps up — FractalBits is designed for it.
AI / ML Training & Checkpointing
Challenge GPUs sit idle when data loaders can't keep up, and non-atomic checkpoint writes risk corrupt restarts.
FractalBits Millions of small reads per second keep accelerators fed, while native atomic rename commits checkpoints and swaps dataset versions instantly — reachable over S3 or POSIX.
Data Analytics & Lakehouse
Challenge Query engines over Parquet and Iceberg-style tables fire huge numbers of small ranged GETs — object-store latency becomes the bottleneck.
FractalBits ~3 ms P99 on 4 KiB objects with deep concurrency cuts scan times, and atomic directory rename makes table compaction and commits safe.
Metadata-Heavy & Small-Object Workloads
Challenge Billions of tiny objects and deep prefixes overwhelm inode-based metadata with heavy distributed transactions.
FractalBits The full-path Fractal ART metadata engine does single-pass lookups — millions of operations per second from a single metadata node, with real directory semantics.
Cost-Efficient S3 Express Alternative
Challenge S3 Express One Zone delivers low latency, but at high cost and locked to AWS.
FractalBits Match or beat that performance at a fraction of the cost — over 100× lower cost per small-object write — deployed into your own AWS or GCP account in minutes, with no EBS.
Benchmarks
Measured, Not Promised
Real numbers from a real cloud deployment — and you can reproduce every one of them with a BYOC deployment of your own.
GET Workload
PUT Workload
All of the above is served by a single metadata node holding hundreds of millions of objects — a direct measure of the Fractal ART engine's efficiency. Need more? Scale out horizontally by adding metadata and data nodes.
Benchmark Configuration
Resources
Documentation
Everything you need to get started and understand how FractalBits works.
See the performance for yourself
Deploy FractalBits to your AWS or GCP account in under 5 minutes and run your own benchmarks.
Contact us: founders@fractalbits.com