NVIDIA AI Accelerators

Unleash AI, HPC, and Visualization with NVIDIA H200, B200 & RTX PRO™ Blackwell GPUs

Best For

• Memory-Bound AI Workloads
• Double-Precision Scientific Simulations (FP64)
• AI-Enhanced HPC Workloads
• Inference at Scale
  

• Large Language Model (LLM) Training
• Multimodal Generative AI
• Multi-GPU Distributed Training
• AI-Accelerated Scientific Research
  

• AI inference at the edge
• Content creation and rendering
• Multi-application workflows in small workstations
• Video encoding/decoding for professional media
  

• Generative AI and neural rendering
• Real-time 3D design and simulation
• Scientific visualization and data science
• High-performance video editing and streaming 
  

• Engineering simulations and VR environments
• Cinematic rendering and neural graphics
• Scientific computing and data visualization
  

GPU

H200 (PCIe, HGX SXM5)

B200 (PCIe, HGX SXM6)

RTX PRO 2000 (PCIe)

RTX PRO 4000 (PCIe)

RTX PRO 4500 (PCIe)

Memory

141GB HBM3e

Up to 192GB HBM3e

16GB GDDR7

24GB GDDR7

32GB GDDR7

AI Performance

Very High

Extreme

Low

Moderate

Moderate

HPC Performance

High

Low

Very High

Low

Low

NVIDIA Tensor Core GPUs 

H200 NVL (PCIe)

FEATURE

• memory:
• 141 GB HBM3e memory 4.8 TB/s bandwidth
  
• FP8 TFLOPS*
• 3341
• FP32 TFLOPS*
• 60
• FP64 TFLOPS*
• 30
• Interconnect
• PCIe Gen 5, optional 2-way and 4-way NVLink bridges 
• MIGs 
• Up to 7 per GPU
• Configurable tdp
• Up to 600W
  

H200 HGX (SXM5)

FEATURE

• memory:
• 141 GB HBM3e memory 4.8 TB/s bandwidth 
• FP8 TFLOPS*
• 3958
• FP32 TFLOPS*
• 67
• FP64 TFLOPS*
• 34
• Interconnect
• 4-way and 8-way NVSwitch
• MIGs 
• Up to 7 per GPU
• Configurable tdp
• Up to 700W
  

NVIDIA H200: Two Paths to AI and HPC Acceleration

The NVIDIA H200 Tensor Core GPU, based on the Hopper architecture, comes in two primary form factors: the H200 NVL and the H200 SXM, which is the module used in the HGX H200 platform. While both variants feature the groundbreaking 141GB of HBM3e memory and its impressive 4.8 TB/s bandwidth, they are engineered for different data center environments.



The H200 SXM (used in the HGX platform) is designed for maximum performance and scalability with its high-bandwidth NVLink interconnect. In contrast, the H200 NVL is a more versatile, lower-power PCIe-based option. The NVL still allows for high-speed GPU-to-GPU communication via 2-way or 4-way NVLink bridges, bridging the gap between PCIe flexibility and NVLink performance.

NVIDIA B200: Powering the AI Revolution

The NVIDIA B200 GPU, built on the cutting-edge Blackwell architecture, represents a significant leap forward in AI and high-performance computing. The B200 is engineered to accelerate training, fine-tuning, and inference for large-scale models, including generative AI and large language models (LLMs). At its core, the B200 features a dual-die design, packing an astonishing 208 billion transistors. This chiplet architecture enables extreme compute density and energy efficiency, making it ideal for training and deploying massive AI models such as large language models (LLMs), generative AI, and scientific simulations.

The B200 features 192 GB of HBM3e memory connected via an 8192-bit interface, delivering up to 8 TB/s of memory bandwidth—eliminating bottlenecks and enabling faster data throughput for memory-intensive tasks. With support for FP4 precision and sparsity, the B200 achieves up to 20 PFLOPS of peak performance, offering up to 5× the inference throughput of its H100 predecessor. Importantly, the B200 is available exclusively in the SXM form factor, designed for high-throughput server environments and optimized for integration into our Procyon and Altezza HGX™ servers.

NVIDIA RTX PRO™ Blackwell GPUs

Redefining Professional Graphics and AI

The NVIDIA RTX PRO Blackwell series marks a transformative leap in professional GPU technology, engineered to meet the demands of the most advanced AI, design, and visualization workflows. Built on the cutting-edge Blackwell architecture, these GPUs deliver unprecedented performance, memory capacity, and efficiency, empowering professionals across industries to push the boundaries of creativity, simulation, and computation.

At the heart of the RTX PRO Blackwell lineup are innovations like 5th-generation Tensor Cores, offering up to 3× the AI performance of previous generations, and 4th-generation RT Cores, which double ray tracing throughput for photorealistic rendering. The new Streaming Multiprocessors (SMs) feature enhanced processing throughput and introduce Neural Shaders, integrating neural networks directly into programmable shaders to drive the next era of AI-augmented graphics.

With support for FP4 precision, DLSS 4 multi-frame generation, and Mega Geometry, RTX PRO Blackwell GPUs enable up to 100× more ray-traced triangles, revolutionizing how professionals create immersive 3D environments and simulate complex systems. The inclusion of GDDR7 memory significantly boosts bandwidth and capacity, with flagship models like the RTX PRO 6000 offering up to 96GB of VRAM, ideal for handling massive datasets and multi-application workflows.

GPU Selection Guide

Why B200?

NVIDIA RTX PRO Blackwell – Professional Visualization & AI

Partner with Atipa Technologies

Model

RTX PRO 2000

RTX PRO 4000

RTX PRO 4500

RTX PRO 5000

RTX PRO 6000 Max-Q

RTX PRO 6000 Server

Memory

16GB GDDR7

24GB GDDR7

36GB GDDR7

48GB GDDR7

96GB GDDR7

96GB GDDR7 ECC

Bandwidth

288 GB/s

~576 GB/s

~864 GB/s

1.34 TB/s

1.79 TB/s

1.79 TB/s

TDP

70W

140W

200W

300W

250W

600W

Form Factor

PCLe, Dual-shot, SFF

PCLe, Single-shot

PCLe, Dual-shot

PCLe, Dual-shot

PCLe, Dual-shot

PCLe, Dual-shot