GIGABYTE AI Server G894-SD3 8U - NVIDIA HGX B300 / 8x SXM GPU / 2048GB RAM / 2x 7,68TB SSD

New arrival 16,760,000,-

GIGABYTE AI Server G894-SD3 8U - NVIDIA HGX B300 / 8x SXM GPU / 2048GB RAM / 2x 7,68TB SSD

ℹ️

This product is intended solely for businesses and educational institutions. Completing an order on the e-shop does not constitute a binding contract, but rather the submission of an enquiry.

The sale of this product is subject to export control, sanctions and other legal restrictions. Before an order can be accepted, the buyer, the end user, the country of delivery and the intended use must be verified. An order can only be accepted after successful authorisation by the supplier.

If the required conditions are not met, the enquiry may be declined or the order cancelled without any entitlement to delivery of the goods.

If you are interested in purchasing an AI solution or accelerator, you can also email us at [email protected]. We will then prepare a proposal and an individual contract for you. For further information, please refer to our terms and conditions (Articles 3.33 and 3.34).

GIGABYTE AI SERVER FOR INFERENCE AND FINE-TUNING OF LARGE AI MODELS

The highest class of AI infrastructure in an 8U format. The GIGABYTE G894-SD3-AAX7 combines the NVIDIA HGX B300 platform with eight Blackwell Ultra GPUs and a pair of Intel Xeon 6776P processors for inference and fine-tuning of large AI models, and as a distributed training compute node.

  • 8× NVIDIA Blackwell Ultra GPUs (HGX B300) — 2.3TB of HBM3e memory (8× 288GB) and NVLink interconnect with up to 1.8TB/s bandwidth per GPU, for an aggregate of 14.4TB/s in an eight-GPU system. For inference and fine-tuning of large models and as a distributed training compute node.
  • 2× Intel Xeon 6776P processors — 64 cores and 350W TDP per processor, 32 slots for DDR5 RDIMM memory (8 channels per processor).
  • Top-tier network connectivity — 8× OSFP with support for 800Gb/s InfiniBand XDR or 2× 400Gb/s Ethernet via NVIDIA ConnectX-8 SuperNIC for distributed AI clusters.
  • Enterprise storage — equipped with 2× 7.68TB NVMe Gen5 SSDs, a total of 8 hot-swap bays for 2.5" NVMe Gen5 SSDs, and 2× M.2 slots for datasets and model checkpoints.
  • Custom build — this configuration has 2,048GB of RAM and 2× 7.68TB SSDs; we can prepare a different memory and disk capacity upon request.
  • Redundant power supply — 12× 3000W power supplies with 80 PLUS Titanium certification; the level of redundancy depends on the configuration.
8×
Blackwell Ultra GPUs
2.3TB
of HBM3e GPU memory
1.8TB/s
NVLink per GPU
2TB
DDR5 RDIMM (32× 64GB)

BUILD BREAKDOWN

Component Details

NVIDIA HGX B300 with eight Blackwell Ultra GPUs
Eight Blackwell Ultra GPUs (B300 SXM) with 2.3TB of HBM3e memory (8× 288GB) are interconnected by NVLink and NVLink Switch technologies, delivering up to 1.8TB/s of bandwidth per GPU (14.4TB/s total). Together, they form a unified compute unit for the inference and fine-tuning of large generative models. The modular GPU board allows for servicing without needing to dismantle the entire chassis.
2× Intel Xeon 6776P
The server is equipped with a pair of Intel Xeon 6776P processors with a 350W TDP on an LGA 4710 socket. Each processor has 8 DDR5 memory channels: RDIMM modules run at up to 6400 MT/s with one module per channel and 6000 MT/s with two, while faster MRDIMMs (up to 8000 MT/s) can only be used with one module per channel. This ensures sufficient bandwidth for data pre-processing and task management across all eight GPUs.
NVIDIA ConnectX-8 SuperNIC
Eight OSFP ports supporting 800Gb/s InfiniBand XDR or 2× 400Gb/s Ethernet via the NVIDIA ConnectX-8 SuperNIC ensure high-speed, low-latency interconnectivity between servers in an AI cluster, which is crucial for distributed training across multiple nodes. Four PCIe Gen5 x16 (FHHL) slots are available for further expansion.
8 hot-swap bays for NVMe Gen5
Eight front hot-swap bays for 2.5" NVMe Gen5 SSDs, equipped with 2× 7.68TB drives in this build, and two internal M.2 slots (PCIe Gen5 x4 and x2) for the operating system or cache. Drives can be replaced without server downtime.

COMPONENT

GPU Cluster

NVIDIA HGX B300

Inference and fine-tuning of large models require more than just powerful GPUs—they need an architecture that unites them into a single whole. The NVIDIA HGX B300 platform in the GIGABYTE G894-SD3-AAX7 server interconnects eight Blackwell Ultra GPUs with NVLink and NVLink Switch technologies, achieving a bandwidth of up to 1.8TB/s per GPU. Meanwhile, the modular design of the GPU board allows for easy servicing without having to dismantle the entire chassis.
The high-speed NVLink interconnect eliminates the bottleneck of a standard PCIe bus when exchanging data between GPUs. As a result, eight GPUs with 2.3TB of HBM3e memory (8× 288GB) behave as a single large accelerator, which is key for the inference and fine-tuning of large language models, multimodal networks, and scientific simulations. NVIDIA ConnectX-8 SuperNIC network cards are integrated directly into the GPU platform, so even communication between servers in a cluster occurs with minimal latency via eight OSFP ports with up to 800Gb/s InfiniBand XDR throughput. When deployed as a compute node for distributed training, where tens to hundreds of GPUs work together on a single model, this interconnect architecture is just as important as the computational power itself.
GIGABYTE AI Server G894-SD3-AAX7
GIGABYTE AI Server G894-SD3-AAX7

COMPONENT

Professional CPU

2× Intel Xeon 6776P

Even the most powerful GPU cluster needs a processor that can prepare and deliver data fast enough. The GIGABYTE G894-SD3-AAX7 server is equipped with a pair of Intel Xeon 6776P Granite Rapids generation processors with 64 cores, 128 threads, and a 350W TDP on an LGA 4710 socket, complemented by 32 DDR5 memory slots with eight channels per processor.
A base frequency of 2.3GHz with a turbo boost of up to 3.9GHz and 336MB of shared L3 cache provide sufficient performance for data pre-processing, task orchestration across the eight GPUs, and I/O management, preventing the CPU from becoming a bottleneck. RDIMM memory modules run at up to 6400 MT/s with one module per channel (6000 MT/s with two), while MRDIMM modules run at up to 8000 MT/s with one module per channel. The dual Xeon 6776P setup thus acts as the system's control layer, ensuring that the eight NVIDIA HGX B300 GPUs process data smoothly and without downtime caused by waiting for the CPU.

Custom Configuration

GIGABYTE AI Server G894-SD3-AAX7

This configuration is equipped with a pair of Intel Xeon 6776P processors, an NVIDIA HGX B300 platform with eight Blackwell Ultra GPUs including ConnectX-8 SuperNIC network cards, 2,048GB of DDR5 RAM (32× 64GB), and 2× 7.68TB NVMe Gen5 SSDs. We can prepare a different memory and storage capacity for you on request: up to 32 DDR5 RDIMM modules (or 16 MRDIMM modules), up to 8 NVMe Gen5 SSDs in hot-swap bays, and 2 M.2 drives. The server is designed for large-scale inference, fine-tuning of large language models, as a computing node for distributed training, and for demanding HPC and scientific simulations.

GIGABYTE AI Server G894-SD3-AAX7

Key Technologies

What makes the GIGABYTE G894-SD3-AAX7 exceptional?

2× Intel Xeon 6776P

350W TDP, 8 channels of DDR5 per processor.

8× Blackwell Ultra GPUs

NVIDIA HGX B300 platform with 2.3TB of HBM3e (8× 288GB).

1.8TB/s NVLink per GPU

Aggregated 14.4TB/s between the eight GPUs.

2TB RAM

32× 64GB DDR5 RDIMM, 6000MT/s

8× OSFP Ports

800Gb/s InfiniBand XDR or 2× 400Gb/s Ethernet via ConnectX-8.

2× 7.68TB NVMe Gen5 hot-swap

Equipped with 2.5" NVMe Gen5 drives, 8 bays in total

12× 3000W Power Supplies

80 PLUS Titanium, redundant.

NVIDIA BlueField-3 DPU Support

Offloads network and security tasks from the main CPU.

Connectivity

Ports and Slots

Front I/O

  • 2× USB 3.2 Gen1 Type-A
  • 1× VGA
  • 2× 10Gb/s RJ45 (Intel X710-AT2)
  • 1× RJ45 management LAN

Rear I/O

  • 8× OSFP (800Gb/s InfiniBand XDR or 2× 400Gb/s Ethernet)
  • 1× RJ45 management LAN (MLAN board)

Dimensions

8U Rack Form Factor

The server occupies 8U in a standard 19" rack. With a depth of 923mm and twelve 3000W power supplies, we recommend checking the rack depth and the power and cooling capacity in your data centre before installation. The maximum power consumption depends on the configuration – we will state the value for your setup in our offer.

GIGABYTE AI Server G894-SD3-AAX7
351mm
height
447mm
width
923mm
depth

Specifications

Technical Overview

ModelGIGABYTE G894-SD3-AAX7
CPU2× Intel Xeon 6776P (64 cores, 350W TDP)
GraphicsNVIDIA HGX B300, 8× GPU Blackwell Ultra SXM
GPU Memory2,304GB HBM3e (8× 288GB)
GPU InterconnectNVLink 5, up to 1.8TB/s per GPU, 14.4TB/s total
Memory2,048GB DDR5 RDIMM (32× 64GB, 6000MT/s), 32 slots, 8 channels per CPU
Installed Drives2× 7.68TB NVMe Gen5 SSD (2.5")
Drive Bays8× 2.5" NVMe Gen5 hot-swap
M.22× (PCIe Gen5 x4 and x2)
Expansion Slots4× PCIe Gen5 x16 (FHHL)
GPU Network8× OSFP, NVIDIA ConnectX-8 SuperNIC (800Gb/s InfiniBand XDR or 2× 400Gb/s Ethernet)
LAN2× 10Gb/s, Intel X710-AT2
ManagementBMC ASPEED AST2600, TPM 2.0 (optional module)
Power Supply12× 3000W 80 PLUS Titanium, redundant (redundancy level depends on configuration)
Form Factor8U rack, 447 × 351 × 923mm (W × H × D)

Warranty and Service

Manufacturer Service Level: GIGABYTE Bronze (Standard RMA)

What's included in the manufacturer service level supplied with this server

  • Service request response: within 3 working days
  • RMA repair: within 14 working days
  • Technical support (helpdesk): 5 days a week, 8 hours a day

Outside the Scope of Bronze Level

What this level does not include

On-site service, advance parts replacement and a dedicated technical manager are not included in this service level. We are happy to offer you higher levels with faster response and on-site service as part of a quote.

Request a Quote for GIGABYTE G894-SD3-AAX7

This server is available by quote only. Tell us your intended deployment, required configuration and timeline, and we will get back to you with an offer. This offer is intended for businesses and institutions.

Nebo nám napište přímo na [email protected]

Děkujeme! Vaši poptávku jsme odeslali, ozveme se co nejdříve.

Questions

Frequently Asked Questions

How many GPUs does the server contain?
Eight NVIDIA Blackwell Ultra GPUs on an HGX B300 platform with 2.3TB of HBM3e memory (8× 288GB), interconnected via NVLink and NVLink Switch with a throughput of up to 1.8TB/s per GPU.
What_s in the build and can I modify it?
This build includes a pair of Intel Xeon 6776P processors, an NVIDIA HGX B300 platform with eight GPUs, 2,048GB of DDR5 RAM and 2× 7.68TB NVMe SSDs. We can prepare a custom quote for a different memory capacity (up to 32 DDR5 modules) and drives (up to 8× NVMe Gen5 and 2× M.2).
What network throughput does the server offer for clustering?
Eight OSFP ports support either 800Gb/s InfiniBand XDR or 2× 400Gb/s Ethernet via the NVIDIA ConnectX-8 SuperNIC.
What memory does the server support?
DDR5 RDIMM up to 6400MT/s with one module per channel (6000MT/s with two) or MRDIMM up to 8000MT/s, but only with one module per channel, i.e. a maximum of 16 modules. There are 32 DIMM slots, 8 channels per processor.
What storage can be installed in the server?
Eight hot-swap bays for 2.5" NVMe Gen5 SSDs in the front panel and two internal M.2 slots (PCIe Gen5 x4 and x2) for the operating system or cache.
What power supply does the server require?
The server has 12 × 3000W power supplies with 80 PLUS Titanium certification. Their sum (36kW) is not the power consumption – the power supplies are oversized for redundancy. The maximum power consumption depends on the configuration and will be specified in the offer; please verify the power and cooling capacity of your rack before installation.
How do I request a server?
We supply the server upon request. Please fill in the form above or email us at [email protected] – we will prepare a configuration, a price and a delivery date for you.

Specifications

Processor

Processor series Intel Xeon
Processor model number Intel Xeon 6700 Series
Max TDP 350 W

Motherboard

Socket LGA 4710
Slots for additional cards PCI Express x16
Design DIMM

disk

Storage Type SSD
Storage capacity (total) 15.36 TB
SSD Capacity 15.36 TB
Internal interface M.2 (PCIe 5.0 4x NVMe)
Max. number of disks 8 pc(s)
Disk size 2.5"
Hot Plug Hot Plug

Colour and design

Colour Grey
Front panel location front
Sidewalls Opaque

PSU

PSU 36,000 W

Memory

Size of operational RAM 2,048 GB
Memory type DDR5
Number of RAM slots 32 pc(s)
Number of installed slots 32 pc(s)

Outputs

Graphics VGA D-SUB
Other LAN

Case

Width 447 mm
Height 351 mm
Depth 923 mm

NCNR (Non-Cancellable, Non-Returnable)

Způsob nákupu pro firmy Na poptávku s individuální smlouvou

GPU configuration

Number of GPUs in configuration 8 ×
NVLink bandwidth 1,800 GB/s

Propojení GPU

Typ propojení SXM, NVLink
Code:  NRn2015g
Warranty: 24 months
Links: Producer's Website:
Manufacturer information

SIL Servis Partner a.s., Těšínská 1970/56, 710 00 Slezská Ostrava, Czech Republic, +420 552 301 141, [email protected]

Gigabyte, founded in 1986, is an irreplaceable player in the field of computer components. They put great emphasis on key technologies and the strictest quality standards. The company produces a wide range of basic computer components such as motherboards and graphics cards, but also have many computer peripherals, laptops, tablets, desktops and even network solutions and servers on offer. Recently they also introduced a line of Gigabyte mobile phones.

Show more