Summarize this news article with AI:
Ask ChatGPT
Ask Grok
Ask Perplexity
Ask Claude
Ask Google AIHigh-performance object storage, a unified view across storage assets, and native vector hosting help enterprises keep GPUs utilized, cut data movement and bring AI into production
Hong Kong, November 5 – Sangfor Technologies, a global provider of cybersecurity, cloud computing, and IT infrastructure solutions, today announced the release of aStor All-Flash 5.3.2R1, the latest version of its all-flash unified file, object and block storage. Built for the AI era, the release introduces high-performance object storage, Data Flow 2.0 for unified data access and version control, and Vector Bucket for native vector storage, together with enhanced disaster recovery and platform capabilities.
As AI moves from experimentation into enterprise production, storage must do more than hold and protect data. AI workloads are reading and moving data at an unprecedented scale, making performance as important as cost efficiency. At the same time, enterprise data remains fragmentd across NAS, object, SAN and legacy platforms, forcing teams to spend valuable time moving and preparing datasets. Meanwhile, the growing adoption of AI agents and RAG applications is driving demand for vector data infrastructure, often requiring organizations to deploy and manage separate vector databases.
Built for this new reality, Sangfor aStor All-Flash 5.3.2R1 combines high-performance object storage, unified data management and native vector hosting on a single platform.
High-Performance Object Storage: Faster from Data Discovery to the GPU
aStor All-Flash 5.3.2R1 rebuilds the object data path for AI workloads such as training data loading, checkpointing, RAG retrieval and Lakehouse analytics:
- Directory Bucket adds a hierarchical namespace to S3-compatible storage. AI and lakehouse engines can locate datasets by directory and narrow LIST operations to a parent directory, while renaming or moving a directory becomes a metadata operation rather than object-by-object copy and delete.
- A rebuilt object engine separates metadata from object data and orchestrates requests by workload, so small-object concurrency and large-object bandwidth can improve together under mixed AI workloads.
- S3 over RDMA keeps S3 semantics while moving object data by RDMA into compute-node memory, bypassing the kernel TCP path. It automatically falls back to TCP when RDMA is unavailable.
- GPU Direct Storage for object, through NVIDIA cuObject, sends object data directly into GPU memory on NVIDIA platforms, removing the host-buffer relay so GPUs spend less time waiting for data.
- Object/file protocol interoperability lets one copy of data be accessed through both file and object protocols, reducing the duplicate copies created only by protocol differences.
Data Flow 2.0: Use Data Where It Already Is
Data Flow 2.0 gives AI teams and agents a unified view and entry point across aStor Hybrid Flash, NAS, object and cloud storage. It shifts the question from “how do we move data to the AI platform?” to “how do we use data where it already is?”
- Metadata first, data on demand. Existing data is onboarded in place: directories and metadata are synchronized into a unified namespace, original data stays at its source, and hot data is cached on the all-flash tier on first access or by proactive pre-warming. Workloads can start without waiting for a full migration.
- Find data by conditions. Users can query across silos by project, time, type, size, tags and source, narrow results into task-specific datasets, and access them through the Sangfor platform, OpenAPI, EDS Skill or Iceberg REST Catalog.
- Real-time change notification. File and object changes can trigger Webhook, Kafka or AMQP notifications, so AI data pipelines and knowledge bases refresh as source data changes rather than days later.
- Git-style data versioning. Dataset-level commit, branch, merge and rollback let teams run experiments and training in isolated, zero-copy branches, trace every version and recover quickly if a release has problems.
- Agent-driven operations (AIOps). Agents can manage storage operations through a unified interface, enabling self-service storage management.
Vector Bucket: Native Vector Hosting on the Storage Platform
Vector Bucket brings vector storage and search into aStor, accessed through an S3 Vector-compatible API. Vectors, indexes and original data are persisted together on the same platform, so enterprises can host block, file, object and vector data in one place instead of building a separate vector database.
- Low memory footprint at scale. A tiered index keeps only navigation information in memory and reads the full index and vectors from all-flash media on demand, while low-bit quantization compresses the navigation index to help maintain retrieval accuracy at lower memory cost.
- Start converged, scale independently. The vector service can start on the same converged nodes with no separate cluster, then scale compute and storage independently as workloads grow.
- Resilient by design. Compute nodes keep only cache and temporary state; vectors, indexes and original objects are persisted in aStor, so a compute-node failure does not put vector data at risk and unfinished tasks are taken over automatically.
Typical uses include enterprise knowledge retrieval (RAG) with results traceable to source documents, shared long-term memory for AI agents, and personalized recommendation.
Enterprise-Grade Protection and Operations
The release also strengthens data protection and platform capabilities for production environments:
- File replication: snapshot-based asynchronous replication for NFS/CIFS workloads, with an RPO as low as 5 minutes, metadata failover RTO of 30 seconds or less, one-click planned switchover and bandwidth throttling.
- Object replication: replication-group-based asynchronous disaster recovery for S3-compatible object storage, with periodic, scheduled and continuous sync strategies to match each workload’s recovery objective. Continuous sync targets RPO in seconds.
- Platform enhancements: QoS for file and object, Fibre Channel support for block storage, and daylight saving time support.
“AI success is no longer determined only by GPUs or models; it depends on how effectively enterprises manage, organize and deliver data,” said Keith Lee, Cloud Business Director of IMD, Sangfor Technologies. “With aStor All-Flash 5.3.2R1, customers can keep GPUs fed with data, put existing data to work without large-scale migration, and host vectors alongside their source data, all on one unified platform.”
Built for Three Core AI Scenarios
- AI Data Lake Foundation. A unified data view over existing NAS and object storage, combined with Directory Bucket, vector storage and data versioning, gives Spark, Trino, DuckDB and AI applications one foundation without moving all data first.
- High-performance AI Training. GDS for both object and file, fast checkpoint reads and writes, and hot, warm and cold auto-tiering balance training throughput with cost.
- Enterprise Knowledge and Agent Memory. Vector Bucket provides shared knowledge and long-term memory for digital employees, intelligent customer service and other agents, without a dedicated vector database for every AI project.
Availability
aStor All-Flash 5.3.2R1 is now available as a beta release. Enterprises and partners interested in evaluating the solution can contact their local Sangfor representatives to request a proof of concept (POC).
Ready to Build Your AI-Native Data Foundation?
Discover how aStor All-Flash 5.3.2R1 helps organizations accelerate AI workloads, unify data management across silos, and simplify infrastructure for AI training, RAG applications, and intelligent agents. Contact your local Sangfor representative to schedule a personalized demo or proof of concept (POC).
Learn more: Sangfor aStor All-Flash Unified Storage (EDS)
Request a consultation: www.sangfor.com/about-us/contact-us
About Sangfor Technologies
Sangfor Technologies is a global leader in cybersecurity, cloud computing, and IT infrastructure, providing fully integrated and AI-driven solutions. Founded in 2000 and publicly listed since 2018 (STOCK CODE: 300454.SZ), Sangfor serves over 100,000 customers worldwide. With over 7,000 employees and more than 70 branch offices across APAC, EMEA, and LATAM, the company is committed to delivering on its mission to Make Your Digital and AI Transformation Simpler and More Secure. For more information, visit www.sangfor.com.
