to the list of articles
to the list of articles
back
back
December 6, 2024
Event
Vacancy
Article
Blackwell Architecture: A New Standard for GPUs in the Age of Artificial Intelligence
The Blackwell microarchitecture, developed by NVIDIA, promises to revolutionize the world of high-performance computing. As AI and machine learning become an integral part of our lives, the new chips provide the necessary power for new developments and the growth of artificial intelligence. What does this mean for the industry, consumers, and the future?
Why is everyone talking about Blackwell?
The Blackwell architecture is a new microarchitecture for GPUs from NVIDIA, succeeding Hopper and Ada Lovelace. The name "Blackwell" honors the American mathematician David Blackwell, whose research in probability theory and information inspired innovative solutions in AI algorithms. The first leaks about the new architecture appeared in 2022, and the official announcement was made on March 18, 2024, at the Graphics Technology Conference (GTC), where NVIDIA CEO Jensen Huang stated that Blackwell is the processor for the era of generative AI and the engine of the new industrial revolution. During the presentation, NVIDIA unveiled two key products: the B100 and B200 accelerators. The Blackwell B200 GPU is already being hailed as "the most powerful in the world." With a performance of up to 20 petaflops FP4 and 208 billion transistors, this chip promises to revolutionize AI computing. For comparison, its predecessor, the H100, offered "only" 4 petaflops.
Blackwell builds on the success of the previous Hopper architecture, which became popular in 2023. In November 2024, Morgan Stanley reported that the entire production volume of Blackwell chips for 2025 had already sold out, highlighting the high demand for the new technology.
Compared to its predecessor, Hopper, the new architecture offers:
– Higher performance: The B100 and B200 accelerators deliver up to 75% faster speeds.
– Optimized energy use: Reduced power consumption allows for more efficient scaling of computing power.
– Integration with the Grace CPU: The combination of the Blackwell GPU with Grace processors creates a unified platform for working with complex AI models.
In addition, NVIDIA's shift to an annual product release cycle allows them to respond more quickly to market challenges. Blackwell is already supported by key industry players: Google, Meta, Microsoft, OpenAI, and Oracle.
Blackwell is aimed at several key sectors:
1. Data Centers: High-performance computing required for training generative models like GPT or DALL-E becomes more accessible with the new GPUs.
2. Cloud Platforms: Cloud service providers will be able to offer clients greater computing power without a significant increase in costs.
3. Scientific Research: Tasks requiring the processing of large datasets, such as simulations or genome analysis, are accelerated.
4. Corporations and Startups: Blackwell enables even small companies to deploy powerful AI solutions, lowering the entry barrier to technology.
This architecture creates new opportunities for AI in business, science, and industry. The manufacturer notes that Blackwell is already being used by all major global cloud service providers, AI innovators, system and server suppliers, regional cloud service providers, and telecommunications companies worldwide.
The new GPU will serve as the foundation for so-called "AI factories"—state-of-the-art data centers specifically designed for generative AI. These centers have already been given many names: supercomputers, server farms, or giga-factories, but they all share one characteristic—unprecedented energy consumption.
Blackwell is aimed at several key sectors:
1. Data Centers: High-performance computing required for training generative models like GPT or DALL-E becomes more accessible with the new GPUs.
2. Cloud Platforms: Cloud service providers will be able to offer clients greater computing power without a significant increase in costs.
3. Scientific Research: Tasks requiring the processing of large datasets, such as simulations or genome analysis, are accelerated.
4. Corporations and Startups: Blackwell enables even small companies to deploy powerful AI solutions, lowering the entry barrier to technology.
This architecture creates new opportunities for AI in business, science, and industry. The manufacturer notes that Blackwell is already being used by all major global cloud service providers, AI innovators, system and server suppliers, regional cloud service providers, and telecommunications companies worldwide.
The new GPU will serve as the foundation for so-called "AI factories"—state-of-the-art data centers specifically designed for generative AI. These centers have already been given many names: supercomputers, server farms, or giga-factories, but they all share one characteristic—unprecedented energy consumption.
New Technologies = New Challenges
The Blackwell development is not just a technical breakthrough, but also a challenge for competitors. NVIDIA's graphics processors have long maintained their leadership in the AI field, and the new architecture only strengthens this influence. Competitors such as AMD and Intel are now in a difficult situation: to remain competitive, they will have to rethink their technological approaches.
At the same time, the demand for servers focused on generative AI is growing at an incredible pace. According to Dell'Oro, sales of such servers will exceed $147 billion in 2025, and new performance standards will contribute to a reduction in computing costs. This will open new opportunities for startups and small companies to integrate innovations into their business processes.
However, Blackwell also faces several manufacturing challenges that complicate the mass deployment of this revolutionary chip. The main issue lies in the overloading of Taiwan Semiconductor Manufacturing Co. (TSMC) production lines, which remains NVIDIA's key partner. The demand for high-tech graphics processors, produced using the latest technologies, far exceeds the current capabilities of factories.
To overcome this, TSMC and NVIDIA are seeking alternative solutions. One of these is relocating part of the production to TSMC's new Arizona plant, which is funded by the US government with $6.6 billion. However, even here, things are not straightforward. Even if the chips are manufactured in the US, they still need to be sent back to Taiwan for the complex packaging process, which extends the production cycle.
Another issue is the environment. Protests have already erupted in Memphis over the construction of Elon Musk's giga-factory, which will use 200,000 GPUs and require massive amounts of energy. Similar challenges are arising in other countries as well. For instance, the Netherlands is restricting the construction of data centers due to the insufficient power capacity of the energy grid.
To successfully implement Blackwell, several key tasks need to be addressed: stabilizing the supply of components and developing environmentally friendly cooling systems for new data centers.
What’s next?
According to forecasts from the International Energy Agency, the demand for electricity to power new data centers could more than double by 2026. This will require innovative approaches to creating energy-saving technologies, and these developments may shape the future.
By mid-2025, NVIDIA plans to stun the market with the new generation Blackwell Ultra GB300 servers. A true breakthrough is expected: increased performance and fully liquid cooling for maximum effect.
What does this mean for the market? Firstly, a significant leap in server performance, which promises to surpass current models. Secondly, higher energy efficiency, albeit at a higher cost. This transition also opens opportunities for companies that specialize in cooling systems and components.
To understand where such technological innovations are steering our world, it’s worth noting that in addition to Blackwell, NVIDIA presented two more projects at the beginning of 2024: GR00T — a universal base model for humanoid robots, and the NIM platform — a solution for rapid deployment of AI models. Each new development pushes us closer to an era where AI and high-performance computing will become part of daily life. Who knows, perhaps our generation will witness how assistant robots and personalized AI systems change the way we work, learn, and even relax.
Author: Irina Zhdanyuk
No items found.
We are waiting for you at the event!
Чекаємо вас на заході!
Oops! Something went wrong while submitting the form.
