Gigabyte G894-SD3-AAX7: 8U Dual-Socket GPU Server for AI & HPC

Afrin Patni
Gigabyte G894-SD3-AAX7: 8U Dual-Socket GPU Server for AI & HPC

In modern AI training, high-throughput compute tasks, and high-performance computing (HPC) environments, infrastructure needs to deliver both high GPU density and balanced CPU resources. The Gigabyte G894-SD3-AAX7 is designed to meet these needs with its 8U form factor, dual Intel Xeon CPUs, and support for eight NVIDIA GPUs via the HGX B300 platform. This makes it an excellent choice for data centers, research environments, and AI labs.

Key Features & Specifications of the Gigabyte G894-SD3-AAX7

Processor & Architecture

  • Supports dual Intel Xeon 6700- or 6500 Series core processors, with up to 350 W TDP per CPU.

  • Uses LGA 4710 sockets and a system-on-chip design for efficient board layouts.

GPU & Acceleration Support

  • Equipped for NVIDIA HGX B300, hosting 8 x Blackwell Ultra GPUs for heavy-duty AI training or HPC.

  • Provides 1.8 TB/s GPU-to-GPU bandwidth using NVLink and NVSwitch for tight communication between GPUs.

  • On-board ConnectX-8 SuperNICs deliver 8 × 800 Gb/s OSFP InfiniBand or dual 400 Gb/s Ethernet for GPU networking.

Memory & Capacity

  • The server supports up to 32 DIMM slots (16 per CPU) of DDR5 memory.

  • Supports both RDIMM and MRDIMM modules; RDIMMs can run up to 6400 MT/s (1DPC) / 5200 MT/s (2DPC), MRDIMMs up to 8000 MT/s (in certain configurations)

  • Total memory capacity can go very high, allowing data-intensive workloads to run smoothly.

Networking & I/O

  • 2 × 10 Gb/s LAN ports via Intel X710-AT2 for general connectivity

  • Rear I/O includes the GPU networking via the ConnectX-8 cards.

  • The management network has a dedicated management LAN port (1 Gbps) for server management.

Storage Options

  • Front-facing 8 × 2.5″ hot-swap NVMe bays (Gen5), ideal for high-speed local data, scratch space, or model datasets.

  • Internal M.2 slots: supports 1 × Gen5 ×4 and 1 × Gen5 ×2 for boot drives or lightweight OS storage.

Expansion

  • Four FHHL PCIe Gen 5 ×16 slots via a bridge board, allowing for additional accelerators, DPUs, or high-speed networking cards.

  • Compatible with NVIDIA BlueField-3 DPUs, making it future-ready for network offloading and secure compute.

Power & Cooling

  • Uses twelve 3000W 80 PLUS Titanium power supplies in redundant configuration to ensure reliability under heavy load.

  • Designed for steady airflow, the server includes multiple fan modules across different zones (motherboard, GPU sleds, PCIe) to maintain thermal stability.

Security & Management

  • Embedded ASPEED AST2600 for video and baseboard management.

  • TPM header for TPM 2.0 to support hardware-based security.

  • Two-ROM architecture to help with firmware resiliency.

Performance & Use-Case Scenarios

Thanks to its architecture and high GPU density, the G894-SD3-AAX7 is well-suited for:

  • Large-Scale AI Training: With 8 Blackwell GPUs and NVLink/NVSwitch interconnect, it can support training of large neural networks and generative models.

  • Inference Clusters: Its high GPU interconnect and memory capacity make it good for inference workloads that require low latency and high throughput.

  • HPC Applications: Scientific simulations, physics modeling, and data-parallel HPC tasks benefit from its dual-CPU setup and fast GPU communication.

  • Data Center AI Infrastructure: Corporations and service providers building AI compute clusters can use this server as a node in a high-density, efficient architecture.

  • DPU-Enabled Architectures: With support for BlueField-3 DPUs, it can offload networking and security tasks, enabling smart, composable data center deployments.

Efficiency & Scalability

  • The use of Titanium-rated redundant power supplies reduces the total cost of ownership by increasing energy efficiency and ensuring system reliability.

  • High-bandwidth GPU interconnect minimizes training time bottlenecks and helps scale compute horizontally.

  • The support for BlueField-3 DPUs gives flexibility to offload tasks like security, storage, or networking, allowing the CPU and GPU to focus purely on compute.

Why Choose the Gigabyte G894-SD3-AAX7?

  • Full 8-GPU Capacity: Built for dense AI compute with HGX B300 support.

  • High Memory Bandwidth: DDR5 with 32 DIMMs supports data-intensive applications.

  • Future-Ready Networking: Built-in SuperNICs for 800 Gb/s InfiniBand or 400 Gb/s Ethernet.

  • Robust Power Design: Twelve 3 kW Titanium PSUs ensure redundancy and stable performance.

  • Modular & Expandable: PCIe Gen5 slots plus DPU support for future upgrades.

  • Secure & Manageable: ASPEED BMC, TPM, and dual ROM architecture help safeguard system integrity.

Availability via Saitech Inc

You can configure and purchase the Gigabyte G894-SD3-AAX7 through Saitech Inc. They offer tailored build options, including a choice of memory, GPU count, storage, and networking modules, helping you design a server that fits your exact AI, HPC, or data center needs.