Move less data. Keep AI compute fed.
Jitronix is a programmable-storage runtime and simulator for validating safe, bounded eBPF processing closer to NVMe storage—before committing to a controller integration.
Verified PoC
AI dataset quality filter
Input dataset
640 KB
10,000 records
Returned
64 KB
1,000 matches
Device-to-host transfer
90% less
Measured in the Jitronix software PoC using synthetic fixed-size records and 10% selectivity. This is a data-reduction result—not a hardware latency or GPU-throughput claim.
Working software PoC
Byte-for-byte verified output
Seeking AI infrastructure design partners
Why now
AI compute is expensive. Moving irrelevant data makes it worse.
Jitronix explores a simple question: when data must be inspected or transformed before use, can a safe program near storage return only what the next stage needs?
Proof, not a promise
A working software PoC with a clear evidence boundary.
The current demonstrator executes an eBPF filter in a sandboxed storage-memory model, returns matching records, and independently compares every output byte with a host implementation.
640,000
input bytes
64,004
device-to-host bytes
1,000
matching records
Verified
against host output
1. Read in sandbox
10,000 synthetic records
2. Execute eBPF
Bounded quality-score filter
3. Return matches
90% less transfer
The platform
A portable execution layer for programmable storage.
Jitronix gives infrastructure teams a controlled way to test whether moving a small, bounded function toward storage is worthwhile—before taking on firmware, silicon, or fleet risk.
Safe eBPF execution
Validate programs and constrain memory access before logic reaches a storage target.
Reproducible workload evaluation
Compare host and storage-side paths using the same input and byte-for-byte output checks.
Hardware-ready architecture
A modular backend designed to progress from software simulation toward NVMe controller integrations.
Development path
Simulator → workload evidence → controller target
Built toward the emerging NVMe Computational Programs model, with hardware validation as the next milestone.
Candidate workloads
Where storage-side execution could earn its place.
The right workload is selective, bounded, and cheaper to evaluate near the data than after moving every byte. We are validating that boundary with infrastructure and controller teams.
Design partners
Bring the workload. We will test the thesis together.
We are looking for AI infrastructure, storage, and controller teams willing to pressure-test whether a real workload belongs closer to the data.
Technical fit review
Map where storage-side processing fits—and where it does not—in your data path.
Workload-specific PoC
Evaluate a bounded filter or transformation using representative data and success criteria.
Reproducible evidence
Compare correctness and data movement now, then define the hardware benchmark that matters.
Start with a 30-minute technical review.
No platform rollout and no hardware commitment. We will begin with your data path, selectivity, record shape, and current bottleneck.