Multi-node GPU Cluster

Multiple GPUs for Large-scale, High-performance AI computing

Multi-node GPU Cluster is a service that provides a physical GPU server without virtualization with the goal of supporting large-scale, high-performance AI computing.
Bare Metal Server with GPU supports the clustering of multiple GPUs. Users can access GPU servers with ease by integrating with high-performance storage and networking products on Samsung Cloud Platform.

Overview

  • Easy-to-use GPU Architecture

    Multi-node GPU Cluster uses Bare Metal Server, which is embedded with highperformance NVIDIA SuperPOD architecture. Using the GPU, it handles multiple user jobs or high-performance distributed workload of large-scale AI model learning.

  • Integration with High-performance Network

    Through integration with the network resources of Samsung Cloud Platform, Multinode GPU Cluster can handle high-performance AI jobs. By configuring GPU direct RDMA (Remote Direct Memory Address) using InfiniBand switch, it directly processes data IO between GPU memories, enabling high-speed AI/Machine learning computation.

  • Integration with High-performance Storage

    Multi-node GPU Cluster supports integration with various storage resources on the Samsung Cloud Platform. A high-performance SSD NAS File Storage is available directly integrated with high-speed networks, and integration with Block Storage and Object Storage is possible.

01

04

Service Architecture

Users - Internet - Multi-node GPU Cluster Multi-node GPU Cluster
Bare Metal Server GPU GPU GPU GPU GPU CPU
NVSwitch
900GB/s (H100)
600GB/s (A100)
InfiniBand Switch HCA ← GPU Direct RDMA Zone → HCA (Right Bare Metal Server)
Bare Metal Server CPU GPU GPU GPU GPU GPU
NVSwitch
900GB/s (H100)
600GB/s (A100)
Block Storage(BM) : Ethernet 25Gbps AFA NAS Storage (Bare Metal Server 구간) High Performance Storage : Ethernet 100Gbps (Bare Metal Server Zone) AFA NAS Storage (A100, H100) (Bare Metal Server 구간) Object Storage(BM) : Ethernet 25Gbps

Key Features

  • Create/manage GPU Bare Metal Server
    1. Standard GPU Bare Metal Server with 8 NVIDIA GPUs
      ※ Internal NVMe disk, NVIDIA NVSwitch and NVIDIA NVLink
    2. Provide OS standard image of RDMA SW Stack (OS : Ubuntu)
  • High performance processing
    1. Configure GPU direct RDMA environment using InfiniBand switch
    2. Provide high-performance SSD File Storage (A100, H100)
  • Storage and network integration
    1. Provide additional storage and network connection (Block, Object) on top of an OS disk
    2. Integration setting for subnet/IP and VPC Firewall
    • Billing
    • Usage-based : Hourly billing based on the time allocated after requesting for resources
    • Commitment-based : Different discount rates applied based on contract years (1 year or 3 years). Monthly rates based on resource type
      ※ Penalty is charged for early termination within the contract period
Let’s talk

Whether you’re looking for a specific business solution or just need some questions answered, we’re here to help

Get in Touch
Share