| |
| AI workloads have two critical stages: |
| • |
Model training, where algorithms learn from vast datasets to recognize patterns, |
| • |
Inference, where trained models apply these patterns to make predictions. |
|
| |
| The storage requirements for these stages differ primarily in scale, performance demands, and data access patterns. |
| |
| Today, let’s take a look at what storage is more suitable for AI training. |
|
| |
Use Case Scenario |
| A research institute conducts AI model training to solve complex problems, enabling it to make significant contributions across various fields. To develop effective AI models, these training workloads leverage vast and diverse datasets and require high-performance storage capable of delivering a read throughput of 140GB/s. This high-performance storage is crucial to accelerate AI training by efficiently managing the intensive, random read/write patterns typical of AI workloads and ensuring that compute resources, such as GPUs, maintain a continuous data flow without delays. |
| |
| Also, the storage solution must be compatible with the Lustre high-performance parallel file system employed by the institute. Moreover, a capacity of 4PB for cold data storage is required to store the large volumes of data used for training. |
| |
|
|
| |
 |
Recommended Product - GS 5000U |
| EonStor GS 5000U series, a high-performance 2U 24-bay U.2 NVMe SSD unified storage solution that can achieve 50GB/s throughput, with 1.3M IOPS and 0.3 ms latency—ideal for data-intensive AI training. |
| |
|
|
|
| |
|
| |
Configuration |
| 1. |
Model: GS 5000U x 3 + JB 3090 x 3 |
| 2. |
Interface: 100GbE x 4 (per appliance) |
| 3. |
Drive: 15.68TB U.2 NVMe SSD x 72 (for GS 5000U) + 18TB HDD x 270 (for JB 3090) |
|
| |
Product Advantages |
| 1. |
Ultra-High Performance: GS 5000U supports 100GbE NVMe over Fabric (NVMe-oF) technology to enable efficient connectivity and boost throughput. In configuration with three appliances, the system reaches 140GB/s performance required for this application scenario. |
| 2. |
High Capacity via JBOD: GS 5000U supports capacity expansion. For this case, the JB 3090, a 4U 90-bay JBOD, was selected. Three appliances provide up to 4,860TB of raw capacity—ideal for storing extensive AI training datasets. |
|
| |
|
|
| |
|
|
| |
Get Started |
| Make an inquiry to Infortrend sales team for more information |
| |
|
|
| |
|
|
|
|
