Nous acceptons la cryptomonnaie — les cryptos que vous détenez peuvent ainsi acheter du matériel réel.Payez en crypto, dépensez-la en matériel réel. Livraison assurée offerte partout dans le monde au-delà de 399 $. Cartons neutres, sans marque, expédiés depuis l’Allemagne.Livraison assurée offerte au-delà de 399 $. Votre taux est bloqué pendant 30 minutes dès l’ouverture de la caisse.Taux bloqué 30 minutes. Chaque article est scellé, son numéro de série est vérifié, et il est couvert par une garantie de 24 mois.Scellé, garanti 24 mois.

Articles suivis Se connecter
5U rack server chassis with eight GPU modules visible

GIGABYTE

GIGABYTE G593-SD2-H200 rack server 8-GPU H200 32 DIMMs

SKU CH-S5FT-15QEK MPN G593-SD2-H200

5U rack server with eight NVIDIA HGX H200 GPUs delivering 900GB/s GPU-to-GPU bandwidth, dual Xeon CPUs and 32 DDR5 DIMM slots

  • 8-GPU HGX H200 platform for large-scale AI training
  • 900GB/s GPU-to-GPU bandwidth via NVLink and NVSwitch
  • Dual Xeon processors for balanced CPU and accelerator performance
  • 32 DDR5 RDIMM slots for high-capacity memory configurations
  • Dual 10Gb/s LAN ports with Intel X710-AT2 controller

High-density AI compute platform

The GIGABYTE G593-SD2-H200 is a 5U dual-processor server built around an NVIDIA HGX H200 8-GPU baseboard delivering 900GB/s GPU-to-GPU bandwidth through NVLink and NVSwitch. It targets organisations running large-scale training or inference workloads that demand maximum accelerator density in a single chassis.

Thermals and power profile

Liquid-cooling infrastructure handles the thermal load of eight H200 GPUs and dual Xeon Scalable CPUs within the 5U envelope. Redundant high-efficiency power supplies feed the full configuration while acoustic output remains constrained by datacentre airflow standards rather than office limits.

Expansion and management extras

Dual ROM architecture provides firmware redundancy for unattended operation. Two 10Gb/s LAN ports driven by an Intel X710-AT2 controller handle management and data traffic. The platform accepts NVIDIA BlueField-3 DPUs or SuperNICs for accelerated networking and storage offload. Thirty-two DDR5 RDIMM slots across eight channels supply memory capacity for model parameters and dataset caching.

Points forts

  • 8-GPU HGX H200 platform for large-scale AI training
  • 900GB/s GPU-to-GPU bandwidth via NVLink and NVSwitch
  • Dual Xeon processors for balanced CPU and accelerator performance
  • 32 DDR5 RDIMM slots for high-capacity memory configurations
  • Dual 10Gb/s LAN ports with Intel X710-AT2 controller

Caractéristiques

Aucune caractéristique publiée pour cet article pour l’instant.

Questions sur cet article

How do I install the eight HGX H200 GPUs and which step is often missed?

The eight GPUs arrive pre-integrated on the HGX baseboard; the step often missed is fully seating the baseboard into the 5U chassis and torquing all retention screws before powering on.

What maintenance will the 32 DDR5 RDIMM slots and dual ROMs need later?

Plan periodic firmware updates for the dual ROM architecture and replace any DDR5 RDIMM that reports correctable errors during memory diagnostics.

What limit appears first when all eight GPUs run sustained workloads?

The 900 GB/s GPU-to-GPU NVLink bandwidth ceiling is reached before CPU or memory bandwidth becomes the bottleneck.

Également dans ce rayon

Les clients ont aussi comparé

Même rayon, même caisse — huit monnaies et un taux bloqué 30 minutes.