TL;DR
Salient Store is a hardware-software co-design framework that transforms edge storage from a passive "bit bucket" into an active participant in the computational pipeline. By offloading video compression and quantum-safe encryption to Computational Storage Devices (CSDs) and reusing inference features for archival, it reduces communication latency by ~6.2× and data volume by ~6.1×.
What Problem Does This Paper Address?
Continuous learning at the edge—essential for applications like autonomous driving and urban mobility—is drowning in data. A single 1080p camera at 60fps generates over 117 TiB of raw data daily. While researchers have optimized the compute side of edge servers (inference and training), the storage and archival systems have remained a bottleneck.
Archiving this data involves three heavy-duty tasks: compression, encryption, and redundancy (RAID). As shown in the authors' analysis of resource utilization, these tasks can consume significant CPU and DRAM resources, often competing with the primary analytics mission (see the table below).
Furthermore, edge devices are physically vulnerable. Standard encryption like RSA faces a "store now, decrypt later" threat from future quantum computers. Current edge systems lack the overhead to implement quantum-safe cryptography without crippling performance.
Methods
The authors propose a "hybrid storage" architecture that integrates Computational Storage Devices (CSDs)—SSDs equipped with their own FPGAs—alongside traditional HDDs.
Key technical innovations include:
- Neural Codec Reuse: Instead of running a separate, expensive video encoder (like H.264), Salient Store reuses the feature extraction layers of the existing inference model (e.g., MobileNet). It then uses a layered neural codec on the CSD's FPGA to compress these features further.
- Temporal Redundancy via Motion Vectors: The system incorporates motion vectors into the neural latent space, allowing it to encode only the changes between frames rather than entire images, similar to traditional inter-frame compression but within a deep learning framework.
- Accelerated Quantum-Safe Cryptography: The authors implemented a hardware-accelerated version of Lattice-Based Cryptography (LBC). They designed specialized modules—High-Speed Schoolbook Polynomial Multiplication (HSPM) and Pipelined Systolic Dimension Modular Multiplier (SDMM)—to run on the CSD FPGAs, providing quantum resilience with minimal latency.
- Peer-to-Peer Data Flow: The design enables CSDs to communicate directly over the PCIe bus, bypassing the host CPU for data-heavy archival tasks.
Datasets
The framework was evaluated across several diverse, high-volume datasets to test its versatility:
- Autonomous Driving/Vision: Waymo Open Dataset, nuScenes, and Cityscapes.
- 3D Spatial Data: KITTI Vision Benchmark Suite (for 3D point cloud data).
- Audio: CHIME Audio data-set (to demonstrate multi-modal capability).
Key Insights and Ideas
- The "Free" Feature Extractor: The most elegant insight is that in a continuous learning system, the video has already been processed by an inference engine. Salient Store treats those intermediate neural features as a partially compressed representation, avoiding redundant computation.
- Storage as a Compute Node: By moving the "archival pipeline" (compression -> encryption -> RAID) into the storage device itself, the host CPU is freed to focus entirely on real-time analytics and model finetuning.
- Future-Proofing the Edge: The transition to lattice-based cryptography is often dismissed as too slow for edge devices. This paper proves that with proper hardware-software co-design on FPGAs, quantum-safe security is viable today.
Key Results
- Latency & Volume: Salient Store achieved a 6.2× reduction in communication latency and a 6.1× reduction in data volume compared to classical storage approaches.
- Hardware Efficiency: The implementation on CSDs performed 1.99× better than using standard high-end Alveo FPGAs, thanks to the proximity of computation to the data.
- Encryption Speedup: The FPGA-accelerated lattice encryption offered a 3.2× speedup over its software counterpart and even outperformed software-based RSA by 2.5×.
- Cost-Benefit Sweet Spot: The authors identified that an 8:1 ratio of standard SSDs to CSDs provides the most optimal balance between system cost and acceleration benefits.
Limitations and Future Work
- Hardware Cost: CSDs are currently significantly more expensive (~15×) than standard SSDs and ~25× more than HDDs. While the performance gains are high, the initial capital expenditure remains a barrier for large-scale deployments.
- Scaling Contention: The authors noted that as the number of storage nodes increases, network contention and orchestration challenges grow exponentially. Future work is needed to optimize data placement to minimize remote disk accesses over the network.
- Vector Databases: While the paper focuses on raw archival, the authors acknowledge that integrating these features into vector databases for efficient retrieval is a logical next step but was beyond the current scope.
Summary updated