Cornerstone of AI Infrastructure

AI adoption is an irreversible trend, yet the true test remains continuously creating value from data.

The GS 5000U is purpose-built for AI infrastructure, empowering enterprises to achieve exceptional performance with a more cost-efficient upfront investment. It is a storage system designed to unshackle GPU computing power to accelerate AI training and inference. Offering flexible, long-term scalability, the GS 5000U serves as the reliable storage backbone for your AI transformation.

Speed, Scale and the Future

Unleash AI Computing Power

The GS 5000U Series delivers up to 50GB/s read throughput and 1.3 million IOPS. Supporting 200GbE, it ensures ample bandwidth to eliminate GPU starvation, accelerating multi-node training and inference. This shortens the model deployment cycle for high-performance AI.

Fully Utilize GPUs, Accelerate Workflows

Leveraging NVMe-oF and GPUDirect Storage (GDS), the system achieves ultra-low latency, as low as 0.3 ms. Its lightning-fast access eliminates checkpoint write stall and maximize GPU utilization. The GS 5000U delivers breakthrough performance for training, inference, and metadata-intensive workloads.

Continuous Availability

Featuring a fully redundant hardware design (controllers, fans, power supplies), the system minimizes downtime risk. Integration with high-availability (HA) architecture provides proactive site-level failover protection, ensuring zero data loss and near-zero downtime with second-level recovery.

Learn more about HA Cluster

Unmatched Future Scalability

The GS U.2 Series ensures long-term value by handling exponential data growth. It offers wide expansion options, including high-density JBODs and high-performance NVMe JBOFs. Configure resources on demand to support hot-tier training data or cold-tier archiving across every stage of AI deployment.

NVMe-oF AI Storage

Build a high-speed, intelligent, and secure AI data storage foundation that empowers enterprises to complete AI workloads faster and more efficiently.

• Storage System with High-Speed Networking: The GS 5000U utilizes NVMe-oF and 200GbE to deliver up to 50GB/s bandwidth and ultra-low latency of just 0.3ms. This performance completely eliminates GPU starvation, ensuring ample throughput for critical hot data and enabling GPUs to operate at maximum efficiency.
• Direct GPU Acceleration: GPUDirect Storage (GDS) ensures data bypasses the CPU and moves directly to GPU memory via the high-speed network. This direct path is critical for eliminating CPU bottlenecks, reducing latency, and maximizing I/O throughput to fully exploit GPU computing power.
• Expansion Enclosures: Infortrend offers enclosures options tailored to the AI data lifecycle: high-speed NVMe SSD JBOFs for scaling hot data, or high-density JBODs for archiving historical assets. This flexible expansion achieves the perfect balance of performance and resource efficiency.

Software Designed for AI Storage

HPC File System Compatibility

The GS File System is POSIX compatible for seamless integration with high-performance parallel file systems like Lustre. This ensures compute servers access data efficiently without bottlenecks, safeguarding data quality and integrity.

Multi-Layer Fault Tolerance Design

The system features built-in RAID fault tolerance, local snapshots, and intelligent algorithms to prevent synchronized SSD failures. This robust storage-level design ensures the highest data security, consistency, and protection with pre-failure alerts.

Cross-Site Recovery Strategy

The system supports synchronous and asynchronous cross-site data replication (e.g., Remote Replication Service) and cloud integration (EonCloud Gateway Service). This guarantees complete data availability for sub-second recovery time objectives (RTO) following a disaster.

Learn more about Local/Remote replication

Learn more about EonCloud Gateway

AI Data Smart Lifecycle Management

Supporting automatic tiering across up to four media layers, intelligent algorithms autonomously handle data migration. A single solution simplifies management, meeting all needs from hot training to cold archiving, maximizing efficiency and cutting cost.

Learn more about Automated Storage Tiering

The Highest-Performing Hybrid Flash U.2 NVMe SSD Unified Storage

EonStor GS 5000U delivers ultimate performance for demanding AI applications. With support for NVMe-oF, 100 / 200 GbE, and 5th Gen Intel® Xeon® Processor, it achieves up to 50GB/s throughput performance and 1.3M IOPS with latency as low as 0.3 ms.

Explore EonStor GS 5000U

Keystone of AI Data and Performance

Different types of AI workloads impose drastically different I/O patterns and performance demands on storage systems:

• Base Model Training: Requires the highest level of sustained throughput to support multi-node, multi-GPU environments. The system must also withstand massive I/O bursts caused by frequent checkpoint writes during the training process.
• Inference / RAG Services: Extremely latency-sensitive, demanding ultra-low latency and exceptionally high IOPS for random I/O. This ensures that under heavy concurrent user requests, models can still be loaded instantly and respond in real time.
• Fine-Tuning: Requires stable and efficient random read/write performance to handle frequent access to base models and checkpoints throughout the training process, thereby accelerating the model iteration cycle.

Base Model Training
(LLM Pre-training)

• Domain-Specific Base Model Training (e.g., Finance, Healthcare, Pharmaceuticals)

Inference

• Enterprise Internal Chatbots
• Document Summarization
• Real-Time Image Classification

Retrieval-Augmented Generation
(RAG)

• Enterprise Knowledge Base Q&A System
• AI-Assisted Customer Service Center

Fine-tuning

• Branded AI Assistant
• Production Line Defect Detection
• Lightweight Edge-Deployed Models

Inference

A large manufacturing enterprise implements an AI-assisted visual defect detection system across four high-speed production lines to identify minute flaws and minimize quality deviations. The system must process 30 FPS without frame loss and maintain millisecond-level latency between image capture and robotic sorting. It must also support archiving all images (defective and sampled) for long-term retention and future model retraining.

Our Advantages

• High-Bandwidth Data Throughput: The GS 5000U delivers 50GB/s of sustained read performance, ensuring continuous, lossless image data transmission to the GPU cluster for zero-frame-loss real-time inspection.
• Sub-Millisecond Real-Time Response: With ultra-low latency of 0.3 ms, total system latency is maintained within the millisecond range, enabling instant AI inference results and real-time robotic decision-making.
• High-Performance Archiving: Offering 20GB/s of sustained write throughput, the system rapidly archives defect images and audit data without compromising front-end inference performance.
• Petabyte-Scale Data Lifecycle Management: Through high-density JBOD expansion and automated tiered storage, the system efficiently handles multi-petabyte image data growth, providing a long-term, cost-effective data archiving solution.

Retrieval-Augmented Generation (RAG)

A major telecommunications operator manages a customer service center with over 400 active agents. The company aims to deploy an AI-assisted support system to enhance retrieval efficiency across its 15TB knowledge base, reducing agent workload and improving response accuracy. The core challenge involves supporting concurrent queries from all 400 agents with sub-second response AI, maintaining both real-time performance and reliability under heavy load.

Our Advantages

• Sub-Millisecond Response: With an ultra-low latency of just 0.3 ms, the system effectively eliminates RAG retrieval bottlenecks, ensuring sub-second AI response times.
• Ultra-High Concurrency Support: Delivering 1.3 million IOPS, it effortlessly handles the I/O load generated by concurrent queries form hundreds of customer service agents querying the knowledge base simultaneously.
• Optimized Data Path: NVMe-oF and GDS create a direct, low-latency path to GPUs. This accelerates model loading, ensuring stable sub-second RAG performance even under heavy concurrency.

Fine-tuning (Domain Adaptation)

A financial institution plans to implement an AI model to enhance the precision of its intelligent investment advisory services. During training, the company will utilize a 16-GPU H100 cluster to fine tune a general-purpose language model with 8 billion parameters for domain-specific applications.

To ensure sustained GPU efficiency, the storage system must deliver 130GB/s of throughput for rapid data loading. Additionally, checkpoint files are written every 15 minutes, and this process must be completed quickly to minimize GPU idle periods and maximize computational utilization.

Our Advantages

• High Performance and Low Latency: Utilizing 200GbE NVMe-oF and a scale-out design, the GS 5000U delivers the required 130GB/s throughput with just three nodes. It maintains 0.3 ms latency, eliminating I/O bottlenecks.
• Direct GPU Data Path: GPUDirect Storage provides a CPU-bypass data path, accelerating model loading and checkpoint access, ensuring maximum training efficiency.
• High Write Throughput, Random I/O: The system delivers 20GB/s sustained write throughput and 1.3M IOPS for random operations. This ensures multi-gigabyte data writes complete within seconds, minimizing GPU idle time.
• Outstanding Scalability: Supports up to three NVMe JBOF expansion units, offering petabyte-scale capacity with seamless scalability, allowing both throughput and capacity to grow in parallel.