Menu

GIGABYTE G894-SD3-AAX7 NVIDIA HGX B300 8-GPU AI Server

$429,433.20

GIGABYTE G894-SD3-AAX7 is an 8U dual-socket server platform with NVIDIA HGX B300 and eight SXM GPUs. It provides onboard ConnectX-8 GPU networking, 32 DDR5 RDIMM/MRDIMM slots and eight Gen5 NVMe drive bays. Installed CPUs, RAM, drives and optional components require confirmation for the supplied configuration.

BrandGIGABYTESKUG894-SD3-AAX7CategoryNetworking & AI Infrastructure

Explore Related

Related products, categories and research.

GIGABYTE G894-SD3-AAX7 is an 8U AI and HPC server platform built around NVIDIA HGX B300 with eight SXM GPUs. Dual Intel Xeon 6 sockets handle host processing, while onboard ConnectX-8 SuperNIC connectivity provides the GPU network interfaces. The model combines a switched GPU fabric with front-access storage and expansion for cluster deployments.

The manufacturer identifies this variant as ordering number 6NG894SD3DR000AAX7*. The specifications below describe platform capabilities. The installed processor models, memory capacity, drives, expansion cards and software licences must be identified in the quotation; a supported option is not evidence that it is installed.

GIGABYTE G894-SD3-AAX7 specifications

Component Manufacturer specification
Chassis 8U; 447 × 351 × 923 mm, width × height × depth
GPU module NVIDIA HGX B300 with 8 SXM GPUs
GPU interconnect NVLink and NVLink Switch; 1.8 TB/s GPU-to-GPU bandwidth
Processors Dual Intel Xeon 6700/6500-Series; two LGA 4710 sockets; processor TDP up to 350 W
System memory 32 DDR5 RDIMM/MRDIMM slots; eight memory channels per processor
GPU networking 8 OSFP ports via onboard NVIDIA ConnectX-8 SuperNIC; 800 Gb/s XDR InfiniBand or dual 400 Gb/s Ethernet
Host LAN 2 × 10 Gb/s via Intel X710-AT2
Storage bays 8 × 2.5-inch hot-swap Gen5 NVMe bays
Internal storage 2 M.2 slots: one PCIe Gen5 x4 and one Gen5 x2; 2280/22110 formats
Expansion 4 full-height half-length PCIe x16 slots with Gen5 x16 lanes
Power supplies 12 × 3000 W 80 PLUS Titanium redundant units; output depends on input voltage
Operating environment 10–30°C; 8–80% relative humidity, non-condensing

Eight-GPU compute and workload planning

The HGX B300 server architecture connects the eight accelerators through NVLink and NVLink Switch. This fabric carries communication between GPUs within the system. It is separate from the external Ethernet or InfiniBand links used between servers, so both layers need consideration when planning distributed workloads.

GIGABYTE lists AI training, AI inference and HPC among the platform applications. For a language-model deployment, check model size, numerical precision, batch size and host data loading against the intended software stack. The published interconnect bandwidth describes a hardware capability; it is not a measured token rate or a guarantee of training throughput.

ConnectX-8 SuperNIC and external fabric

Eight rear OSFP ports provide GPU networking through onboard NVIDIA ConnectX-8 SuperNIC hardware. The documented modes are 800 Gb/s XDR InfiniBand or dual 400 Gb/s Ethernet. Choose the fabric mode, switch ports, optics and cables together. An OSFP connector alone does not establish that a particular cable or switch configuration will work.

Two front 10 Gb/s LAN ports use an Intel X710-AT2 controller. The system also has front and rear gigabit management LAN interfaces; when both management ports are cabled, GIGABYTE states that the front interface is the default. Keep management access requirements distinct from the high-bandwidth GPU network design.

The platform is compatible with NVIDIA BlueField-3 DPUs. That compatibility does not establish that a DPU is included in this listing. Confirm the card model and slot allocation if infrastructure offload is part of the project.

DDR5 MRDIMM, host CPUs and local storage

Xeon 6700/6500 processor selection affects memory support and expansion availability. GIGABYTE specifies DDR5 RDIMM operation up to 6400 MT/s at one DIMM per channel, or 6000 MT/s at two DIMMs per channel. DDR5 MRDIMM support reaches 8000 MT/s on selected Xeon 6 processors with P-cores, and only with one DIMM per channel.

These are supported speeds rather than a statement of the installed RAM. Match processor support, DIMM type, population and capacity before choosing a configuration. If only one CPU is installed, some memory and PCIe functions may be unavailable according to the manufacturer.

Eight hot-swap NVMe bays allow local data and scratch storage, while the two internal M.2 connections offer different lane widths. Both M.2 slots originate from CPU 1. Identify boot drives and workload drives separately, including capacity, endurance and any required redundancy. Bay counts and slot counts do not indicate that drives are supplied.

Rack, electrical and cooling requirements

The 923 mm chassis depth needs to be checked against rack clearance, rail positioning and rear cabling space. Its fan-cooled design includes dedicated airflow for the GPU tray, motherboard, OSFP ports and PCIe cards. Facility cooling must accommodate the actual configured system load while maintaining the documented operating temperature range.

The twelve power supplies have voltage-dependent output. GIGABYTE specifies a maximum of 1450 W per supply at 115–127 V AC, 2900 W at 200–220 V AC and 3002.4 W at 220–240 V AC. The nominal 3000 W rating must therefore be considered alongside the available electrical supply. Adding supply ratings together does not establish measured system consumption or usable redundancy.

C19 power cords are required. GIGABYTE explicitly states that power cords are not included in its server packages. Check the final electrical plan, cord selection and supplied accessories for this order before installation.

Configuration and manufacturer documentation

GIGABYTE lists a server, two CPU heatsinks, four CPU carriers and an L-shape rail kit in the manufacturer packaging specification. Confirm the delivered package against the written bill of materials, particularly installed CPUs, memory, drives, optional cards and licences. Unit condition and support terms also require confirmation for the exact offer.

Use the official GIGABYTE specifications and support resources to check the current model documentation and qualified operating-system list. For related equipment, browse Networking & AI Infrastructure. Send the required CPU, RAM, storage and fabric configuration through the configuration enquiry page.

Platform questions

Is this a standalone graphics card?

No. This model is an 8U server platform using an HGX B300 module with eight SXM GPUs. It requires server rack, electrical and cooling infrastructure.

Are all 32 memory slots populated?

The manufacturer specification describes 32 available slots. Installed DIMM quantity and capacity are configuration details that must be confirmed for the offered system.

Does the 1.8 TB/s figure describe the external network?

No. It describes the GPU-to-GPU NVLink fabric. External GPU networking uses the separate OSFP interfaces and their selected InfiniBand or Ethernet mode.

Recently viewed

Add AI Robot Supplier as a preferred source on Google.

WhatsApp Chat with our sales team