Sangfor aStor All-Flash Unified Storage (EDS)

Advanced Features & Value Propositions

Ultimate High Performance

Ease of Use & Intelligence

Enhanced Security & Reliability

AI-Native Data Services

Extreme Throughput: up to 120GB/s read and 40GB/s write from a single all-flash unit, with GPU utilization improved 2–5x for AI workloads on Directory Bucket-compatible datasets.
AI Acceleration, End to End: GPUDirect Storage (GDS) for object bypasses CPU and memory buffers to build a direct data channel between NVMe storage and GPU memory, cutting latency and host-side relay for training and inference.
S3 over RDMA: replaces TCP with RDMA for S3 transport, offloading the data plane from the kernel protocol stack to NIC hardware for higher object throughput and lower latency — with no dependency on the NVIDIA GPU ecosystem.
NVIDIA cuObject Integration: extends object performance from the storage server all the way to GPU memory, removing host-buffer relay for GPU-bound pipelines.

Unified Data View: presents heterogeneous storage (NAS, object, cloud, aStor Hybrid Flash) through a single logical namespace, so AI agents and applications always work with the latest data without directory-by-directory traversal or upfront migration.
Object/File Protocol Interoperability: one copy of data, accessible through both file and object protocols, eliminating duplicate storage and the data movement that protocol differences usually force.
Data Versioning for AI Development: Git-style branch, commit, merge and rollback on the data lake foundation, enabling low-cost data isolation testing, version traceability, reproducible training runs, and rapid production rollback.

Asynchronous Replication for File and Object: snapshot-based async replication for NFS/CIFS file workloads (minute-level RPO, as low as 5 minutes) and replication-group-based async DR for S3-compatible object storage (RPO as low as seconds in continuous-sync mode), each with a full disaster-recovery operations lifecycle — failover, restore, recovery protection, reverse sync, and planned switchover.
FC Support for Block: adds Fibre Channel (FC-SCSI) connectivity for block workloads, extending aStor All-Flash’s protocol coverage alongside existing iSCSI/NVMe-oF support.
Encryption: data encryption capabilities are extended in the new version to help meet enterprise and regulatory data-protection requirements.
Existing protections carried forward: 99.9999% multi-level reliability, AI-driven ransomware protection, WORM compliance, and continuous end-to-end data integrity checks.

Directory Bucket: S3-standard-compatible bucket with a hierarchical namespace, dramatically improving metadata operations and retrieval efficiency for large-scale AI datasets — upgrading object access from “finding objects by key” to “locating datasets by directory.”
Vector Bucket: a native, storage-embedded foundation for hosting vector embeddings and indexes alongside the original unstructured data, reducing cross-system data movement during agent inference and lowering the barrier to building enterprise knowledge bases and RAG applications.
Data Flow 2.0 — Unified View and On-Demand Flow: a metadata-first architecture that gives AI agents a single, always-current view of data across storage assets, with hot data promoted to all-flash on demand and returned to lower-cost tiers when a job finishes — balancing performance and cost automatically.
Object Storage, Reconstructed for AI: beyond Directory Bucket and S3 over RDMA, the object engine adds GPUDirect Storage for object, object/file protocol interoperability, and a Git-inspired Data Versioning Engine — together transforming object storage from a passive capacity tier into an active AI data platform.

Ultimate High Performance

Extreme Throughput: up to 120GB/s read and 40GB/s write from a single all-flash unit, with GPU utilization improved 2–5x for AI workloads on Directory Bucket-compatible datasets.
AI Acceleration, End to End: GPUDirect Storage (GDS) for object bypasses CPU and memory buffers to build a direct data channel between NVMe storage and GPU memory, cutting latency and host-side relay for training and inference.
S3 over RDMA: replaces TCP with RDMA for S3 transport, offloading the data plane from the kernel protocol stack to NIC hardware for higher object throughput and lower latency — with no dependency on the NVIDIA GPU ecosystem.
NVIDIA cuObject Integration: extends object performance from the storage server all the way to GPU memory, removing host-buffer relay for GPU-bound pipelines.

Ease of Use & Intelligence

Unified Data View: presents heterogeneous storage (NAS, object, cloud, aStor Hybrid Flash) through a single logical namespace, so AI agents and applications always work with the latest data without directory-by-directory traversal or upfront migration.
Object/File Protocol Interoperability: one copy of data, accessible through both file and object protocols, eliminating duplicate storage and the data movement that protocol differences usually force.
Data Versioning for AI Development: Git-style branch, commit, merge and rollback on the data lake foundation, enabling low-cost data isolation testing, version traceability, reproducible training runs, and rapid production rollback.

Enhanced Security & Reliability

Asynchronous Replication for File and Object: snapshot-based async replication for NFS/CIFS file workloads (minute-level RPO, as low as 5 minutes) and replication-group-based async DR for S3-compatible object storage (RPO as low as seconds in continuous-sync mode), each with a full disaster-recovery operations lifecycle — failover, restore, recovery protection, reverse sync, and planned switchover.
FC Support for Block: adds Fibre Channel (FC-SCSI) connectivity for block workloads, extending aStor All-Flash’s protocol coverage alongside existing iSCSI/NVMe-oF support.
Encryption: data encryption capabilities are extended in the new version to help meet enterprise and regulatory data-protection requirements.
Existing protections carried forward: 99.9999% multi-level reliability, AI-driven ransomware protection, WORM compliance, and continuous end-to-end data integrity checks.

AI-Native Data Services

Directory Bucket: S3-standard-compatible bucket with a hierarchical namespace, dramatically improving metadata operations and retrieval efficiency for large-scale AI datasets — upgrading object access from “finding objects by key” to “locating datasets by directory.”
Vector Bucket: a native, storage-embedded foundation for hosting vector embeddings and indexes alongside the original unstructured data, reducing cross-system data movement during agent inference and lowering the barrier to building enterprise knowledge bases and RAG applications.
Data Flow 2.0 — Unified View and On-Demand Flow: a metadata-first architecture that gives AI agents a single, always-current view of data across storage assets, with hot data promoted to all-flash on demand and returned to lower-cost tiers when a job finishes — balancing performance and cost automatically.
Object Storage, Reconstructed for AI: beyond Directory Bucket and S3 over RDMA, the object engine adds GPUDirect Storage for object, object/file protocol interoperability, and a Git-inspired Data Versioning Engine — together transforming object storage from a passive capacity tier into an active AI data platform.

Target Use Cases: aStor All Flash

AI Data Training (Autonomous Driving/LLM)

Accelerates the entire AI pipeline. From high-speed checkpoint writing (up to 40GB/s) to rapid data loading for training sets, Sangfor aStor All Flash keeps expensive GPUs fully utilized. Features like data preheating load training sets from cold tiers to flash before jobs start to prevent idle time.

Healthcare (PACS & Pathology)

Delivers the high throughput required for loading massive medical images and pathology slides instantly. It supports mixed workloads of small and large files, ensuring rapid retrieval for diagnostic efficiency while offering cost-effective large-capacity storage for archiving.

High-Tech Electronics Manufacturing (EDA)

EDA Chip Design: Accelerates the chip design cycle by handling the massive concurrency of millions of small files generated during simulations. Its high OPS capabilities prevent storage bottlenecks during intense verification phases.

High-Tech Electronics Manufacturing (AOI)

AOI Quality Inspection: Supports stable, high-throughput writing of uncompressed images (RAW/BMP) from production lines, ensuring zero frame loss. Intelligent data tiering automatically moves older inspection logs and images to lower-cost storage, significantly reducing retention costs.

Entertainment (8K Post-Production)

Provides stable, high-bandwidth streams for 8K video editing and rendering. It supports collaborative non-linear editing (NLE) across Mac/Linux/Windows workstations with consistently high read speeds (exceeding 15GB/s in tested scenarios).

Government (Satellite & Aerial Imaging)

Supports the high-concurrency demands of GIS applications, such as oblique photography and remote sensing. The system handles massive small file access for rapid rendering and processing of terabytes of geospatial data.

AI Data Lakehouse Foundation

Lightweight, high-performance foundation for AI data lakehouses without large-scale data movement. Unified Data View unifies NAS, object, cloud, and aStor Hybrid Flash storage, enabling Trino, Spark, DuckDB, and AI pipelines to access lakehouse tables and multimodal datasets.