Nvidia has announced that the future of AI will be driven by its next generation of GPUs, specifically the Blackwell hardware. During the Nvidia GTC 2024 event, CEO Jensen Huang showcased the new Blackwell chips, promising tremendous strides in AI power and efficiency.
The flagship Blackwell “superchip,” GB200, is scheduled for release later this year. It possesses the remarkable ability to scale up from a single rack to an entire data center, solidifying Nvidia’s position as a frontrunner in the AI race.
Blackwell is a substantial leap forward compared to its predecessor, Hopper. Notably, Blackwell boasts 208 billion transistors across its two GPU dies, a significant increase from Hopper’s 80 billion. These GPU dies are interconnected through a high-speed 10 TB/second chip-to-chip link, creating a unified and powerful GPU. This transformative upgrade allows Blackwell to deliver up to 30 times the AI inference performance of Hopper, offering an exceptional 20 petaflops of FP4 power, surpassing any competing product currently available.
CEO Jensen Huang emphasized not only the remarkable performance gains of Blackwell compared to Hopper but also its size. Huang stated, “Blackwell’s not a chip, it’s the name of a platform. Hopper is fantastic, but we need bigger GPUs.”
Nvidia asserts that Blackwell presents significant cost and energy savings, up to 25 times more efficient than its predecessor. For instance, training a model with 1.8 trillion parameters used to require 8,000 Hopper GPUs and consume 15 megawatts of power. However, with just 2,000 Blackwell GPUs, this can now be accomplished using only four megawatts.
The GB200 combines two Nvidia B200 Tensor Core GPUs with a Grace CPU, creating a massive superchip that propels AI development. It offers seven times the performance and four times the training speed of an H10O-powered system.
Additionally, Nvidia unveiled the next-generation NVLink network switch chip, equipped with an impressive 50 billion transistors. This breakthrough technology enables seamless communication between 576 GPUs, generating an astounding 1.8 terabytes per second of bidirectional bandwidth.
Nvidia has already secured partnerships with major industry players to develop Blackwell-powered systems. AWS, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and other prominent names have all committed to integrating this powerful technology into their offerings.
As the demand for AI continues to grow, Nvidia’s Blackwell superchip is poised to revolutionize the industry, driving progress and innovation in AI development and implementation.
Frequently Asked Questions (FAQ)
What is Blackwell?
Blackwell is Nvidia’s next-generation superchip, comprising powerful GPUs and CPUs. It offers substantial improvements in AI power and efficiency compared to its predecessor, Hopper.
How fast is Blackwell compared to Hopper?
Blackwell is up to 30 times faster than Hopper when it comes to AI inference tasks. It delivers an impressive 20 petaflops of FP4 power.
What are the cost and energy savings with Blackwell?
Nvidia claims that Blackwell can reduce costs and energy consumption by up to 25 times. For instance, training a model with 1.8 trillion parameters previously required 8,000 Hopper GPUs and 15 megawatts of power but can now be achieved with just 2,000 Blackwell GPUs consuming four megawatts.
Which companies are partnering with Nvidia for Blackwell?
Nvidia has partnered with major industry players such as AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure, among others, to develop Blackwell-powered systems.
What is the NVLink network switch chip?
The NVLink network switch chip is a next-generation technology from Nvidia that enables seamless communication between 576 GPUs. It creates 1.8 terabytes per second of bidirectional bandwidth.
Where can I find more information about Blackwell?
For further details, you can visit Nvidia’s official website: https://www.nvidia.com
Nvidia’s announcement of its next-generation GPUs, specifically the Blackwell hardware, has significant implications for the AI industry. The upcoming flagship Blackwell “superchip,” GB200, is expected to be released later this year and is poised to solidify Nvidia’s position as a front runner in the AI race.
Blackwell represents a substantial leap forward compared to its predecessor, Hopper. It boasts a remarkable 208 billion transistors across its two GPU dies, a significant increase from Hopper’s 80 billion. These GPU dies are connected through a high-speed 10 TB/second chip-to-chip link, creating a powerful and unified GPU. This transformative upgrade allows Blackwell to deliver up to 30 times the AI inference performance of Hopper, offering an exceptional 20 petaflops of FP4 power, surpassing any competing product currently available.
One of the notable advantages of Blackwell is its size. CEO Jensen Huang emphasized that Blackwell is not just a chip, but a platform. This distinction reflects the need for larger GPUs to meet the growing demands of AI development and implementation.
In addition to its performance gains, Blackwell also presents significant cost and energy savings. Nvidia claims that it is up to 25 times more efficient than its predecessor. For example, training a model with 1.8 trillion parameters previously required 8,000 Hopper GPUs and 15 megawatts of power. However, with just 2,000 Blackwell GPUs, this can now be accomplished using only four megawatts.
Nvidia’s advancements in AI hardware go beyond the Blackwell superchip. The company also unveiled the next-generation NVLink network switch chip, which enables seamless communication between 576 GPUs, generating an astounding 1.8 terabytes per second of bidirectional bandwidth.
To further demonstrate the relevance and impact of Blackwell, Nvidia has already secured partnerships with major industry players. AWS, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and other prominent names have committed to integrating this powerful technology into their offerings.
As the demand for AI continues to grow, Nvidia’s Blackwell superchip is set to revolutionize the industry, driving progress and innovation in AI development and implementation.
For further details about Blackwell, you can visit Nvidia’s official website: https://www.nvidia.com