The problem
Continuous-learning edge servers swallow huge volumes of video and sensor data. With conventional storage, every byte the model wants to look at has to cross the I/O bus into the host’s DRAM, run through compression and encryption on the CPU, then move again into the accelerator. The data movement, not the compute, is the bottleneck.
What Salient Store does
We push the data-intensive operations into the storage itself. FPGA-equipped computational storage devices (CSDs) handle:
- Neural compression — learned codecs tuned for the workload’s data distribution, not generic gzip.
- Quantum-safe encryption — done at rest, in the device, not in the CPU.
- Redundancy and error correction — co-designed with the codec so we don’t pay for both layers separately.
The host sees a thinner, validated stream and a much lower CPU load. The training loop doesn’t change its API.
Results
Prototype results show 6× faster end-to-end data handling and 6.1× less data movement across the host’s I/O fabric. The win compounds when you scale: many drives, each doing local work, aggregate into a serious bandwidth advantage that pure host scaling can’t match.
Why it matters
Continuous learning at the edge fails silently when the I/O can’t keep up with the model’s appetite. Salient Store moves the bottleneck from a shared resource (host DRAM) to a parallelizable one (per-device FPGAs). That’s the architectural change that makes large-scale on-edge training practical.