NVIDIA Blackwell Explained: The Chip Architecture Powering the Next Wave of AI

Artificial intelligence may be defined by models such as ChatGPT, Claude and Gemini, but behind many of today’s most demanding AI systems is another race taking place: the race to build enough computing power to train and run them. One of the biggest names in that race is NVIDIA Blackwell.

Blackwell is NVIDIA’s GPU architecture created for a new generation of artificial intelligence, accelerated computing and high-performance workloads. It succeeds the Hopper architecture that powered systems built around GPUs such as the NVIDIA H100.

But calling Blackwell simply a new graphics chip misses much of the story. The architecture was designed to make enormous AI workloads faster and more efficient while allowing many GPUs to work together almost like one extremely large computing system. That matters because modern AI is becoming increasingly demanding.

What exactly is NVIDIA Blackwell?

NVIDIA introduced the Blackwell platform in March 2024 as its next major architecture for accelerated computing and generative AI. The name comes from David Harold Blackwell, the influential American mathematician and statistician.

At the heart of the data-center version of Blackwell is an unusually large GPU design containing 208 billion transistors. Instead of manufacturing the GPU as one enormous piece of silicon, NVIDIA uses two large dies connected together through a high-speed chip-to-chip link capable of transferring data at up to 10 terabytes per second.

From the software’s perspective, those two pieces can operate as a unified GPU. This approach allows NVIDIA to build a processor substantially larger and more capable than would normally be practical with a traditional single-die design.

Blackwell was built around the demands of modern AI

AI models require enormous amounts of mathematical computation. Training a large language model involves repeatedly processing huge collections of numbers representing model parameters, activations and data. Running the finished model, known as inference, also requires significant computing power when millions of users are sending requests.

Blackwell introduces a new generation of NVIDIA Tensor Cores specifically designed to accelerate these types of calculations. One particularly important addition is support for lower-precision formats such as FP4.

Lower precision means that AI calculations can sometimes represent numbers using fewer bits. When this can be done while maintaining suitable model quality, the processor can perform more operations while using less memory. For large AI systems, improvements like this can have a major impact on how quickly a model responds and how much infrastructure is required to run it.

The Transformer Engine is a major part of Blackwell

Transformers form the foundation of many modern generative AI systems, including large language models. Blackwell includes NVIDIA’s second-generation Transformer Engine, designed specifically to improve training and inference for transformer-based models.

The engine can dynamically work with different numerical precision formats depending on what a particular calculation requires. This allows the hardware to prioritize efficiency where lower precision is sufficient while retaining greater precision where it is needed.

It may sound like a small technical detail, but at the scale of today’s largest AI systems, reducing the amount of memory and computation required for each operation can translate into significant gains.

One GPU is no longer enough

Perhaps the most important part of Blackwell is not what happens inside a single GPU. It is what happens when hundreds or thousands of GPUs need to work together.

Modern frontier AI models are often too large and computationally demanding for a single processor. Training them requires massive clusters containing large numbers of GPUs, and adding more processors only helps if they can exchange information quickly enough.

That is where NVLink becomes important. Blackwell uses NVIDIA’s fifth-generation NVLink technology to provide extremely high-speed communication between GPUs. Instead of treating every GPU as a mostly independent device, NVIDIA can connect many Blackwell GPUs into much larger computing systems.

GB200 shows how large Blackwell can scale

One of the clearest examples is NVIDIA’s GB200 NVL72. This is not a normal server with one or two graphics processors.

The rack-scale system combines 72 Blackwell GPUs and 36 NVIDIA Grace CPUs, connected through NVIDIA’s NVLink infrastructure. The 72 GPUs can communicate across a large NVLink domain, allowing the system to operate more like one enormous computing platform.

NVIDIA designed systems like this for workloads involving extremely large AI models, including models containing hundreds of billions or even trillions of parameters. A single rack therefore becomes less like an ordinary collection of servers and more like a purpose-built AI supercomputer.

Memory is becoming just as important as processing power

Raw computing performance gets much of the attention when new GPUs are announced, but AI workloads are also heavily dependent on memory. Models have to be stored somewhere while calculations are taking place.

The NVIDIA B200 GPU provides 180 GB of HBM3e memory per GPU, with memory bandwidth reaching up to around 8 TB per second. An eight-GPU HGX B200 system can therefore provide approximately 1.44 TB of GPU memory.

Large amounts of fast memory allow bigger portions of an AI model to remain close to the processors performing the calculations, reducing the need to constantly move information from slower storage. For modern AI infrastructure, memory capacity and bandwidth can be just as important as the number of calculations a GPU can perform.

Blackwell is also about efficiency

The AI industry faces another major challenge: power. Large AI data centers can require enormous amounts of electricity, and simply adding more processors indefinitely is not an ideal solution.

NVIDIA has therefore focused heavily on improving the amount of useful AI work that can be completed for a given amount of energy. Some Blackwell systems also use liquid cooling to manage the heat created by dense collections of high-performance processors.

NVIDIA says its GB200 NVL72 platform can deliver substantially greater performance per unit of power than previous-generation H100-based infrastructure under specific AI workloads. Exact improvements will always depend on the model, workload and system configuration, but efficiency is becoming one of the most important measurements for future AI hardware.

Blackwell also includes dedicated reliability technology

When an AI cluster contains thousands of expensive processors running continuously, reliability becomes extremely important. A small hardware problem can interrupt jobs that may have already been running for days.

Blackwell includes a dedicated Reliability, Availability and Serviceability engine, commonly referred to as the RAS Engine. It is designed to monitor hardware and help identify possible reliability problems before they create larger interruptions.

This is another example of how modern AI chips are evolving beyond simply performing calculations faster. At data-center scale, monitoring, networking, cooling and reliability all become part of the computing platform.

Blackwell is not only for enormous data centers

The Blackwell name has also moved into consumer hardware. NVIDIA’s GeForce RTX 50 Series is based on the RTX Blackwell architecture, bringing parts of the same broader generation of technology to desktop PCs and laptops.

These GPUs include fifth-generation Tensor Cores and fourth-generation ray-tracing cores, along with AI-focused technologies used for graphics and content creation. That means Blackwell now stretches from enormous rack-scale AI systems all the way to PCs used by gamers, developers and creators.

The workloads are very different, and a GeForce RTX card should not be confused with a B200 data-center GPU, but they share the same broader architectural generation.

AI is changing what a GPU is designed to do

Graphics processors were originally built primarily to accelerate graphics. Over time, their highly parallel architecture made them useful for many other workloads, and machine learning transformed that role even further.

Today, a high-end NVIDIA data-center GPU is designed not merely to render images but to train neural networks, run generative AI models, process enormous datasets and communicate with large numbers of other processors.

Blackwell demonstrates how far that transition has progressed. Features such as FP4 processing, Transformer Engines and enormous NVLink systems exist largely because the requirements of AI computing are becoming very different from traditional computing.

Blackwell Ultra pushes the architecture further

Blackwell itself is no longer the newest member of the family. NVIDIA followed it with Blackwell Ultra, an expanded version of the platform designed particularly for AI reasoning, agents and workloads that use additional computation while generating an answer.

Systems based on Blackwell Ultra include the GB300 generation. For example, the GB300 NVL72 combines 72 Blackwell Ultra GPUs with 36 Grace CPUs and dramatically increases the amount of fast GPU memory available in a rack.

This evolution shows how quickly AI hardware is changing. Instead of GPU generations being used mainly to provide higher frame rates or incremental performance gains, new generations are increasingly being designed around changes in how AI models themselves operate.

Why Blackwell matters

Blackwell matters because the progress of artificial intelligence does not depend only on better algorithms. It also depends on whether those algorithms can actually run at a practical scale.

A model can become more capable by growing larger, processing more information or spending more computation reasoning about a request, but every one of those improvements creates additional hardware requirements.

Architectures such as Blackwell attempt to make those workloads possible by combining faster AI processing, enormous memory bandwidth and extremely high-speed communication between GPUs. That can affect everything from how quickly an AI assistant responds to how much it costs a company to operate millions of AI requests.

The bigger picture

The most interesting thing about NVIDIA Blackwell may not be any individual specification. It is the scale at which modern computing is being redesigned around artificial intelligence.

Processors, memory, networking, cooling and software are increasingly being developed together as one large AI platform. The result is something very different from the traditional idea of simply installing a faster chip into an existing computer.

Blackwell systems can occupy entire racks, connect dozens of GPUs and move enormous amounts of information between processors every second. At the same time, parts of the Blackwell architecture have reached ordinary PCs through the GeForce RTX 50 generation.

As AI becomes integrated into more software, creative tools, research systems and everyday devices, the hardware underneath those experiences will become increasingly important. NVIDIA Blackwell is one of the clearest examples of that shift.

The AI people interact with may appear as a simple text box on a screen, but behind it can be an extraordinary amount of specialized computing infrastructure working together.

Webitlio’s Tech Picks to Check Out on Amazon

If you enjoy exploring new technology, accessories and smart devices, here are five products worth checking out on Amazon.

Product pricing, availability and specifications can change over time. Check the current Amazon listing for the latest information before purchasing.