The Next Generation of AI Computing: NVIDIA’s Journey from Hopper to Blackwell

The landscape of AI computing is witnessing a revolutionary transformation with NVIDIA’s latest advancement from Hopper to Blackwell architectures. In March 2025, NVIDIA unveiled the Blackwell Ultra AI factory platform, marking a pivotal shift toward the age of AI reasoning. This platform represents not just an incremental improvement but a transformative leap that enables unprecedented capabilities in AI training, reasoning, and inference.

The journey from Hopper to Blackwell represents a significant leap in technology, designed to meet the evolving demands of artificial intelligence (AI) and high-performance computing (HPC). This blog will delve into the key features of both architectures, highlighting their technical advancements and the broader impact on AI development.

Overview of NVIDIA’s Hopper Architecture

Technical Foundation

The Hopper architecture, named after computing pioneer Grace Hopper, was built on TSMC’s 4nm process technology. It introduced the fourth-generation Tensor Cores and pioneered the Transformer Engine, specifically designed to accelerate AI workloads. The architecture featured NVLink 4.0 interconnect technology and utilized HBM3 memory, setting new standards for high-performance computing.

Key Performance Metrics

The Hopper architecture represents a significant leap in GPU technology, delivering substantial improvements across multiple performance metrics while maintaining focus on data center efficiency. From dramatic AI performance boosts to enhanced memory operations and advanced interconnect capabilities, Hopper establishes itself as a comprehensive solution for modern computing challenges.

Key features of the Hopper architecture:

1. Computational Performance Improvements:

• FP8 format support with 6x FP32 throughput increase
• 3x improvement in FP64 performance
• 30x boost in AI performance

2. Memory and Instruction Optimization:

• Enhanced memory operation latency and throughput
• Optimized tensor core instruction performance

3. Interconnect Technology:

• PCIe Gen 5 support
• Fourth-generation NVLink
• Enhanced interconnectivity capabilities

4. Data Center Characteristics:

• Focus on energy efficiency
• Suitable for large-scale deployment
• Optimized for advanced data centers and AI applications

Major Breakthroughs and Improvements

• Transformer Engine: Enables dynamic precision adjustments, optimizing AI workloads.
• DPX Instructions: Enhance dynamic programming tasks, such as sequence alignment.
• Distributed Shared Memory: Improves multi-GPU communication and scalability.

Blackwell Architecture: What’s New and Improved?

Key technological highlights

The Blackwell architecture, announced at GTC 2024, is designed to revolutionize generative AI and large-scale AI workloads. It features advanced NVLink technology, custom Tensor Cores, and a RAS engine, enabling real-time AI models with up to 10 trillion parameters. Blackwell aims to significantly reduce energy usage for AI inference by up to 25 times.

Highlight significant performance improvements

The new architecture delivers unprecedented performance gains:

• 3x improvement in FP8 performance
• 4x faster AI training capabilities
• 40x boost in inference performance
• Significantly better power efficiency ratios

These improvements enable faster training of larger AI models and more efficient inference operations.

Technical Comparison: Hopper vs. Blackwell

NVIDIA’s transition from Hopper (H100) to Blackwell (B200) architecture marks a significant technological leap in GPU design. The new Blackwell architecture demonstrates substantial improvements across all key metrics, with notable advancements in transistor count, memory bandwidth, interconnect speed, and AI performance.

The following table provides a detailed comparison of specifications between the two generations:

Feature	Hopper (H100)	Blackwell (B200)
Fabrication	TSMC 4N (80B transistors)	TSMC 4NP (208B transistors)
Memory	80 GB HBM3 (3 TB/s)	144 GB HBM3e (4.8 TB/s)
NVLink	4th-gen (900 GB/s)	5th-gen (1.8 TB/s)
AI Performance	4 PFLOPS (FP16)	20 PFLOPS (FP16)
Energy Efficiency	Baseline	25x improvement in inference
Key Use Cases	Broad AI/HPC workloads	Generative AI, trillion-param models

Hopper and Blackwell in Real-World AI Applications

Hopper Architecture Products

The Hopper architecture is available across several product lines:

Product	H100 SXM	H100 PCIe	H200 SXM
Memory	80GB HBM3	80GB HBM2e	141GB HBM3e
Memory Bandwidth	3.35 TB/s	2.04 TB/s	4.8 TB/s
NVLink	900 GB/s	N/A	900 GB/s
Max TDP	Up to 700W	350W	Up to 700W
Multi-Instance GPUs	Up to 7 MIGs @10GB	N/A	Up to 7 MIGs @16.5GB

Blackwell Architecture Products

Blackwell introduces a new generation of AI accelerators:

Product	B100	B200	GB200 (Grace Blackwell)
Design	Blackwell GPU	Blackwell GPU	2× B200 GPUs + Grace CPU
Memory	192GB HBM3e	192GB HBM3e	384GB HBM3e
Memory Bandwidth	8 TB/s	8 TB/s	16 TB/s
NVLink	1.8 TB/s (5th-gen)	1.8 TB/s (5th-gen)	1.8 TB/s (5th-gen)
FP4 Performance	7 PFLOPS	9 PFLOPS	20 PFLOPS
TDP	700W	1000W	N/A

Generative AI model training capabilities

Both architectures show exceptional capabilities in training large AI models, with Blackwell demonstrating up to 3x faster training speeds compared to Hopper. Key improvements include enhanced transformer processing, optimized memory bandwidth, and more efficient parallel processing capabilities. The architectures enable training of increasingly large language models while reducing power consumption and training time.

Blackwell Ultra: The Next Evolution in AI Computing

Announced on March 18, 2025, at GTC 2025, Blackwell Ultra represents the latest evolution of NVIDIA’s AI computing platform.

Technical Specifications and Improvements

• Product line includes GB300 NVL72 and HGX B300 NVL16
• GB300 NVL72 delivers 1.5x more AI performance than GB200 NVL72
• HGX B300 NVL16 compared to Hopper architecture achieves:
  • 11x faster inference
  • 7x more compute
  • 4x larger memory
• Integrates NVIDIA Spectrum-X Ethernet and Quantum-X800 InfiniBand, providing 800Gb/s data throughput per GPU
• Features BlueField-3 DPUs for multi-tenant networking, GPU compute elasticity, and real-time security detection

Advanced Memory and Cooling Solutions

• Utilizes ConnectX-8 SuperNIC for high-performance RDMA capabilities
• Optimized for AI factories and cloud data centers to eliminate performance bottlenecks
• Supports various configurations:
  • 72 Blackwell Ultra GPUs
  • 36 Arm Neoverse-based Grace CPUs
  • Rack-scale design functioning as a single massive GPU

LLM Optimization Features

• Optimized for AI reasoning and agentic AI:
  • Supports reasoning and iterative planning for complex multi-step problems
  • Capable of real-time synthetic video generation for training
• Software support:
  • New open-source NVIDIA Dynamo inference framework
  • Support for NVIDIA Llama Nemotron Reason models
  • Integration with NVIDIA AI Enterprise software platform
• Expected availability through major server manufacturers and cloud service providers in second half of 2025

Choose Novita AI for your Trusted GPU Partner

As AI computing continues to evolve with NVIDIA’s groundbreaking architectures, choosing the right GPU cloud partner becomes crucial. Novita AI offers a comprehensive platform that provides access to high-performance GPUs with significant cost advantages

If you’re interested in Novita AI, follow these steps to get started:

Conclusion

The journey from Hopper to Blackwell represents a significant leap in AI computing capabilities. While Hopper set new standards for AI training and scientific computing, Blackwell pushes these boundaries further with dramatic improvements in performance, efficiency, and versatility. This evolution not only advances NVIDIA’s technical leadership but also accelerates the broader adoption of AI across industries. As we look to the future, these architectural advancements will continue to enable new possibilities in artificial intelligence and scientific computing.

Frequently Asked Questions

Novita AI is an AI cloud platform that offers developers an easy way to deploy AI models using our simple API, while also providing the affordable and reliable GPU cloud for building and scaling.

Recommended Reading

GPU Comparison for AI Modeling: A Comprehensive Guide

How Much RAM is Needed for Machine Learning?

Choosing the Best GPU for Machine Learning in 2025: A Complete Guide