Nvidia’s Blackwell Chips: Powering the Next Generation of AI

Nvidia has officially unveiled its next major leap in computing architecture. Named “Blackwell” after mathematician David Blackwell, this new platform is designed to succeed the incredibly popular Hopper H100 chips. As artificial intelligence models grow from billions to trillions of parameters, the hardware running them must evolve. The Blackwell B200 and GB200 are Nvidia’s answer to this demand, offering massive improvements in speed, efficiency, and scalability for data centers worldwide.

The Architecture: Breaking the Reticle Limit

The most significant engineering change in the Blackwell B200 GPU is its physical design. Previous chips were single dies restricted by the “reticle limit,” which is the maximum size a chip can be manufactured using current lithography technology. Nvidia circumvented this physical barrier with a clever architectural shift.

The Blackwell B200 is effectively two massive dies connected together. These two silicon squares are linked by a 10 terabytes per second (TB/s) chip-to-chip interconnect. This connection is so fast that the software treats the two dies as a single, unified graphics processing unit.

Key Specifications

Transistor Count: The B200 packs 208 billion transistors. For comparison, the previous Hopper H100 contained 80 billion.
Manufacturing Process: It uses a custom-built 4NP TSMC process.
Memory: The architecture supports up to 192GB of HBM3e memory, providing 8 TB/s of memory bandwidth.

This dual-die approach allows Nvidia to pack more compute power into a single package without waiting for the semiconductor manufacturing process to shrink further.

Performance Leaps: The Second-Generation Transformer Engine

The primary purpose of Blackwell is to run Generative AI. To do this efficiently, Nvidia introduced a Second-Generation Transformer Engine. This engine allows the chip to support double the compute and model sizes by using new 4-bit floating point (FP4) precision.

Previous generations relied heavily on 8-bit (FP8) or 16-bit (FP16) precision. By optimizing for FP4, the Blackwell B200 can deliver up to 20 petaflops of AI performance. This is a drastic increase from the 4 petaflops offered by the H100.

This shift is critical for inference, which is the process where the AI actually generates answers or images after it has been trained. The ability to run large language models (LLMs) faster means services like ChatGPT or Gemini can respond quicker and serve more users simultaneously.

The GB200 Superchip and NVL72

While the B200 is the individual GPU, the true power for data centers comes from how these chips are combined. Nvidia introduced the GB200 Grace Blackwell Superchip. This component connects two B200 GPUs with one Grace CPU over a 900 GB/s ultra-low-power NVLink connection.

For enterprise-level data centers, Nvidia scales this up even further with the GB200 NVL72. This is a liquid-cooled rack system that acts as a single massive GPU. It connects 36 Grace CPUs and 72 Blackwell GPUs.

Why the NVL72 Matters

Unified Memory: The system provides 30TB of unified memory. This allows data centers to load models with trillions of parameters into a single memory space without partitioning them inefficiently across different servers.
Communication Speed: The NVLink Switch allows 130TB/s of bandwidth across the rack. Every GPU in the rack can talk to every other GPU at full speed.
Liquid Cooling: The density of these chips generates immense heat. The NVL72 uses liquid cooling tech that runs fluid through the compute trays, eliminating the need for loud, energy-intensive fans.

Energy Efficiency and Cost Reduction

One of the loudest criticisms of the AI boom is the immense energy consumption required to train models. Nvidia CEO Jensen Huang addressed this directly during the Blackwell announcement.

Under the previous Hopper architecture, training a 1.8 trillion parameter model (roughly the scale of GPT-4) would require 8,000 H100 GPUs and consume 15 megawatts of power. With Blackwell, Nvidia claims the same task can be accomplished with just 2,000 B200 GPUs consuming only 4 megawatts.

This represents a 25x reduction in cost and energy consumption for running massive generative AI models. For data center operators like Amazon AWS, Microsoft Azure, and Google Cloud, this efficiency is vital for maintaining profit margins while scaling AI services.

Industry Adoption and Partners

The industry demand for Blackwell was immediate upon its announcement. The major cloud providers have already queued up to integrate these systems into their infrastructure:

Amazon Web Services (AWS): Plans to build a massive server cluster (Project Ceiba) utilizing 20,736 GB200 Superchips.
Google Cloud: Will integrate the GB200 into its new cloud architecture to support internal models and Vertex AI customers.
Microsoft Azure: Is adopting the GB200 to power OpenAI’s future model training.
Oracle: Announced it is using Blackwell to power its sovereign AI cloud solutions.

Conclusion

Nvidia’s Blackwell architecture is not just a standard upgrade; it is a fundamental shift in how data centers are built. By moving to multi-die chips and liquid-cooled racks that function as singular supercomputers, Nvidia is paving the way for AI models that are significantly smarter and larger than what we have today. As these chips enter production in late 2024, the bottleneck for AI development will likely shift from hardware limitations to energy availability.

Frequently Asked Questions

What is the price of the Nvidia B200 GPU? While official pricing varies based on volume and contracts, Nvidia CEO Jensen Huang suggested in interviews that the B200 GPU could cost between $30,000 and $40,000 per unit. Complete systems like the GB200 NVL72 rack cost significantly more, likely reaching into the millions.

When will Blackwell chips be available? Nvidia expects to start shipping Blackwell products later in 2024. Major cloud providers are expected to have instances online for customers by early 2025.

Is Blackwell for gaming? The B200 and GB200 are strictly for data centers and enterprise AI. However, the underlying “Blackwell” architecture will eventually trickle down to consumer graphics cards, likely powering the future GeForce RTX 50-series (RTX 5090 and 5080) for gamers.

How does Blackwell compare to Hopper H100? Blackwell offers roughly 2.5x to 5x the training performance and up to 30x the inference performance of the Hopper H100, depending on the specific model and precision used (specifically FP4). It also consumes significantly less energy per token generated.