NVIDIA HGX B300 GPU Servers

Extreme compute for AI training, inference, and HPC workloads

Accelerate complex AI and data-intensive applications with purpose-built NVIDIA HGX B300 servers. Powered by next-generation NVIDIA Blackwell architecture, these systems deliver unmatched performance for enterprises building LLMs, generative AI, multimodal models, and scientific simulations.

Overview

The NVIDIA HGX B300 platform is engineered for high-density GPU performance, offering up to 16 NVIDIA Blackwell GPUs interconnected with NVIDIA NVLink for breakthrough compute in enterprise AI. Designed for scalability, low latency, and power efficiency, HGX B300 systems power cutting-edge training pipelines and high-throughput inference environments.

Whether you are building large language models, running real-time AI services, or supporting complex research workloads, HGX B300 servers provide a flexible, secure, and reliable infrastructure foundation.

Key Features

Up to 16 NVIDIA Blackwell GPUs (NVL16)

Delivering massive compute density for accelerated AI workloads.

NVIDIA NVLink 5.0

High-bandwidth GPU-to-GPU communication for ultra-scalable training.

NVIDIA Tensor Cores (Latest Generation)

Optimized for FP8, FP16, BF16, and other precision modes.

HGX Baseboard Architecture

Supports demanding HPC and enterprise AI environments.

High-Performance PCIe Gen5 Architecture

Ensures maximum throughput for networking and storage.

Multi-Node Scalability

Ideal for building large GPU clusters and AI data centers

Ideal Use Cases

📘

Large Language Model (LLM) Training

Train massive AI models faster and more efficiently with high-density GPU acceleration.

Generative AI (text, image, video, multimodal)

Create advanced generative AI outputs with optimized performance across multiple data types.

AI Inference at Scale

Run AI inference workloads in real-time for enterprise applications with minimal latency.

🔬

Deep Learning Research

Accelerate experimentation and model development for research and innovation.

🖥️

HPC Workloads

Handle high-performance computing tasks, including simulations, modeling, and scientific computations.

🧪

Scientific Modeling & Simulation

Process complex scientific models quickly using GPU-accelerated parallel computing.

Featured NVIDIA HGX B300 Server Configurations

Choose from validated NVIDIA HGX B300 solutions designed for AI training, inference, and HPC workloads.

G894-ZD3-AAX7

  • Barebone: G894-ZD3-AAX7
  • GPU: 8x B300 288GB HGX GPU
  • CPU: AMD TURIN 9555 (64C, 3.2GHz, 360W)
  • Memory: 96GB DDR5 RDIMM 6400MHz

Starting at $460,000.00

View configuration

AS-8126GS-NB3RT

  • Barebone: AS-8126GS-NB3RT
  • GPU: 8× NVIDIA Blackwell HGX B300 NVL8 GPUs
  • CPU: AMD TURIN 9555 (64C/128T, 3.2GHz, 360W)
  • Memory: 96GB DDR5 RDIMM 6400MHz ECC Registered

Starting at $471,675.00

View configuration

NVIDIA HGX B300 Specifications

Category Specification
Form Factor / GPUs 8× NVIDIA Blackwell Ultra SXM GPUs
Tensor Core Performance FP4: up to 144 PFLOPS peak / 105 PFLOPS sustained
FP8 / FP6: up to 72 PFLOPS
FP16 / BF16: up to 36 PFLOPS
TF32: up to 18 PFLOPS
INT8: up to 2 POPS
Compute Performance FP32: up to 600 TFLOPS
FP64 / FP64 Tensor: up to 10 TFLOPS
Memory Total GPU memory: up to 2.3 TB
GPU Interconnect NVIDIA NVLink 5th Generation with NVSwitch
GPU-to-GPU bandwidth: up to 1.8 TB/s
Total NVLink throughput: up to 14.4 TB/s
Networking & IO Network bandwidth: up to 1.6 TB/s
Optimized for distributed AI training and inference clusters
AI Optimization Up to 2× attention performance improvement for LLMs and agentic AI workloads