Nvidia’s Blackwell : A New Era of Computing
Nvidia has constantly been at the forefront of GPU development ever since they released the first GPU, the GeForce 256, all the way back in 1999. Driven by an undying passion for pushing the boundaries of graphics processing, AI (Artificial Intelligence) platforms, and high-performance computing (HPC), Nvidia has consistently redefined the technological landscape. Their latest innovation, named Blackwell after renowned mathematician David Blackwell (who made significant contributions to game theory, probability theory, information theory, and statistics), promises to revolutionize the industry again.
Blackwell is expected to deliver unprecedented speed, performance, and computing power, specifically in artificial intelligence, to many leading tech-organizations, such as Google, Meta, Microsoft, Amazon, and Dell. Possessing the extraordinary ability to build and run real-time generative AI on trillion-parameter large language models, an unparalleled feat compared to other commercial GPUs, Blackwell promises to bring AI to every sector imaginable, unlocking breakthroughs in data processing, engineering simulation, electronic design automation, computer-aided drug design, quantum computing, and generative AI.
Fortunately, Blackwell has arrived at the perfect time. The demand for large-scale AI computing has never been higher, especially as technology continues to develop endlessly, crossing into the boundaries of science fiction. Industries such as healthcare, EVs (Electric Vehicles), robotics, and finance could be redefined entirely for the better by large-scale AI computing. Furthermore, chatbots such as ChatGPT and AI image-generation such as Pixlr only stand to improve as Blackwell becomes public, increasing the access and accuracy of these services and, in turn, improving our lives as well. NVIDIA has been successful with Hopper and Ampere (previous GPU architecture), and at a high level, Blackwell aims to bring more features, flexibility, and transistors.
To understand precisely how Blackwell will redefine the ‘processing world,’ this article looks into the modern hardware Blackwell boasts.
Blackwell’s Key Hardware Features:
Improved Processing Units
For starters, the Blackwell GPU is going to be big. Much bigger than anything Nvidia has produced in the past. In fact, the B200 modules feature two GPU dies on a single package, each around 800mm2 each.
Within these GPU dies lies a new generation of CUDA cores, the backbone of Nvidia’s GPUs, which handle the majority of general-purpose tasks. Interestingly, the two dies can operate as “one unified CUDA GPU,” improving performance, speed, and parallelism without significant compromises. Compared to the 16,896 FP32 CUDA Cores in Hopper, Nvidia’s previous GPU architecture, it’s safe to assume that Blackwell will carry many more.
These cores are expected to offer higher clock speeds and improved power management, making Blackwell the ideal GPU for large-scale, power-consuming AI computing. The benefit of unifying the CUDA cores is also shown in the high bandwidth I/O links between the two dies, coined NV – High Bandwidth Interface (NV-HBI) by Nvidia. The total NV-HBI of Blackwell is 10TB/second, meaning that each die can send 5 TB/second of data in both directions simultaneously. Again, the speed at which Blackwell will operate exceeds well beyond any commercial two-chip GPU currently available.
Additionally, Blackwell’s GPU features an unreal amount of Tensor Cores, which were first introduced in Nvidia’s Volta. Tensor Cores are specialized GPU cores developed by Nvidia specifically for AI performance, machine learning, and high-performance computing (HPC) tasks. More specifically, Tensor Cores enable mixed-precision computing, dynamically adapting calculations to accelerate throughput while preserving accuracy and providing enhanced security. In all, by packing in as many Tensor Cores as possible, Nvidia has optimized Blackwell for AI-related workloads.
Finally, Blackwell also boasts an impressive number of RT Cores (Ray Tracing), which are ASICs, or application-specific integrated circuits dedicated to performing ray tracing operations with extreme efficiency. This means that artists, game developers, and video editors can use Blackwell to create photorealistic displays, environments, lighting, and shadows, all in real time, improving the quality of life for all users.
Advanced Memory Architecture
Besides processing speeds, memory bandwidth and memory capacity are two extremely critical factors in GPUs, especially for GPUs involved in AI, ML, and HPC applications that often require massive datasets, such as Blackwell. Fortunately, Blackwell is expected to introduce new memory architecture that offers higher bandwidth and lower latency compared to Hopper and Ampere.
For instance, the memory clock in Blackwell is 8Gbps, compared to Hooper’s 5.23Gps and Ampere’s 3.2Gbs. Similarly, the memory bandwidth in Blackwell is 8TB/sec, compared to Hooper’s 3.35TB/sec and Ampere’s 2TB/sec. Moreover, the memory bus width in Blackwell is 2x 4096-bit (2x due to the two dies), compared to Hooper and Ampere’s 5120-bit. [A larger bus width means more data can be transferred, significantly improving performance.] These improvements in memory bandwidth and memory capacity are primarily due to the HBM3E and GDDR7 memory stacks that Blackwell contains.
Each of Blackwell’s dies is paired with 4 stacks of HBM3E memory, a high-performance 3D memory architecture used for HDS (high-demand systems), for a total of 8 stacks and an effective bus width of 8192 bits. The considerable improvement in memory capacity seen in Blackwell is especially promising because the one major constraining factor in AI development over the last few years has been memory capacity due to the vast data sets required to train artificial intelligence. With the improved memory bus width, Blackwell, once again, shines as the optimal tool for companies, businesses, and research facilities looking to harness AI.
Blackwell will also implement GDDR7 (Graphics Double Data Rate) chips, the newest generation of graphics memory cards which blows its predecessor, the GDDR6 and GDDR6X, out of the water. The primary benefit of Blackwell running the GDDR7 is its exceptional speed, which, according to Hardware LUXX, is up to 32Gbps. This means that larger quantities of data can be moved quicker within the CPU, reducing bottlenecks and improving performance. Further, the higher the chip density, which is detonated by the number (7 > 6), the more data the chip can hold. The GDDR7 has been shown to hold up to 64Gb per chip, double that of the GDDR6X, which is currently used in commercial CPUs.
Interconnects
Lastly, Nvidia’s NVLink technology, a wire-based protocol, is expected to be updated and improved in Blackwell’s architecture. NVLink is instrumental in enabling HPC (high-performance computing) and AI development, as it enhances performance by allowing high-speed communication between GPUs, addressing previously established communication limitations. Blackwell features two NVLink 5 (1800 GB/sec), while previous GPUs only future one. Coupled with improvements, Blackwell will offer even faster interconnections between GPUs, which is especially important in multi-GPU setups.
Applications
Now that we’ve seen all the various ways Blackwell has redefined the GPU landscape, it’s time to turn our attention to how Blackwell will redefine our landscape. The answer, unsurprisingly, is in AI. Blackwell was specifically designed for AI and high-performance computing (HPC) applications. Everything from the multitude of enhanced Tensor Cores to the substantial increases in memory bus width and bandwidth were engineered for executing optimal AI models. Following the release of Blackwell, which is set for mid-to-late 2024 for industrial use and early 2025 for commercial use, there’ll be a massive leap in artificial intelligence.
It’s important to recognize that the potential applications of Blackwell extend far beyond simple AI tasks like chatbots or image creation. For starters, Blackwell will allow for faster AI-model training, reducing the cost and time required to create important AI models. In turn, these AI models are extremely valuable in scientific research. Blackwell holds promise in various research fields by speeding up complex simulations and tedious tasks, facilitating breakthrough discoveries in drug research and climate modeling.
Furthermore, another potential application of Blackwell is in autonomous vehicles, where Blackwell can power complex decision-making algorithms in real time, resulting in safer and more efficient transportation. Blackwell will also be instrumental in healthcare, providing faster and more accurate diagnosis through advanced personalized medical imaging systems. Blackwell will help optimize large-scale trading algorithms in finance, leading to better investment strategies and more efficient markets. In conclusion, the power and speed of Blackwell opens up a myriad of AI technologies in nearly every sector imaginable, ultimately leading to a better tomorrow.
Written by By: Rocky Klopfenstein
Works Cited
“Nvidia Blackwell Platform Arrives to Power a New Era of Computing.” NVIDIA Newsroom, 18 June 2024,
nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-co mputing
Phillips, Gavin. “GDDR6 vs. GDDR6X vs. GDDR7: What’s the Difference?” MUO, 2 May 2024,
Smith, Ryan. “Nvidia Blackwell Architecture and B200/B100 Accelerators Announced: Going Bigger with Smaller Data.” RSS, AnandTech, 18 Mar. 2024,
