 |
| ANNOUNCING NVIDIA BLACKWELL PLATFORM FOR TRILLION-PARAMETER SCALE GENERATIVE AI |
In a momentous event, NVIDIA CEO Jensen Huang revealed the company's latest innovation poised to revolutionize accelerated computing: Blackwell. Standing before an eager audience, Huang emphasized the critical need for larger GPUs to propel the field forward. To meet this challenge head-on, NVIDIA introduced the Blackwell platform, named in honor of the esteemed mathematician David Harold Blackwell, renowned for his contributions to game theory and statistics.
 |
| Pascal vs Volta vs AMpere vs Hopper vs Blackwell |
The Blackwell architecture, succeeding the groundbreaking NVIDIA Hopper, promises unprecedented performance enhancements. Huang proudly showcased the Blackwell chip, a behemoth in comparison to its predecessor, boasting 2.5 times the performance for training in FP8 and an astounding 5 times in FP4 for inference per chip. Central to its capabilities is the fifth-generation NVLink interconnect, delivering double the speed of Hopper and supporting scaling up to an impressive 576 GPUs.
At the heart of this innovation lies the NVIDIA GB200 Grace Blackwell Superchip, seamlessly connecting two Blackwell NVIDIA B200 Tensor Core GPUs to the NVIDIA Grace CPU via an ultra-low-power NVLink chip-to-chip interconnect, operating at a staggering speed of 900GB/s.
Huang unveiled a groundbreaking system, embodying the unparalleled potential of Blackwell. This compact marvel, boasting memory coherence, marks a paradigm shift where immense computing power converges seamlessly, fostering collaboration akin to a "big happy family" working cohesively on complex applications.
For optimal AI performance, GB200-powered systems can harness the prowess of the NVIDIA Quantum-X800 InfiniBand and Spectrum-X800 Ethernet platforms, delivering cutting-edge networking capabilities at speeds reaching up to 800Gb/s.
The implications of Blackwell extend beyond mere performance enhancements. NVIDIA introduces the NVLink Switch, a pivotal component facilitating the scaling of Blackwell. Each switch can connect four NVLink interconnects at an astounding 1.8 terabytes per second, streamlining communication and optimizing efficiency through in-network reduction.
Enter the NVIDIA GB200 NVL72, a monumental leap in computational prowess, encapsulating the essence of "one giant GPU." This liquid-cooled, rack-scale system harnesses the might of Blackwell, offering unparalleled compute power with 720 petaflops of AI training performance and 1.4 exaflops of AI inference performance in a single rack.
 |
| Database Join query throughput comparing GB200 NVL72, 72x H100, and 72 x86 CPUs |
 |
| Cadence SpectreX simulator runs 13x quicker on a GB200 than on an x86 CPU |
 |
| The GB200 delivers 30x real-time throughput compared to the H100 |
 |
| Cadence Fidelity, a large eddy simulator for CFD applications, runs simulations up to 22x faster on GB200 than x86 CPU |
Further pushing the boundaries, NVIDIA unveils the DGX SuperPOD, an AI supercomputer powered by NVIDIA GB200 Grace Blackwell Superchips. This highly efficient, liquid-cooled rack-scale architecture boasts 11.5 exaflops of AI supercomputing prowess at FP4 precision, coupled with 240 terabytes of fast memory, with the potential for scalability with additional racks.
 |
| NVIDIA DGX SuperPOD with DGX GB200 Systems Al infrastructure with constant uptime. |
As Huang aptly puts it, "data centers are going to be thought of as AI factories," ushering in an era where intelligence generation is paramount.
The industry's response to Blackwell has been resounding, with endorsements pouring in from luminaries such as Sundar Pichai, Andy Jassy, Demis Hassabis, Mark Zuckerberg, Satya Nadella, Sam Altman, Larry Ellison, and Elon Musk.
Blackwell's adoption spans across major global cloud service providers, pioneering AI companies, system and server vendors, and regional cloud service providers and telcos worldwide, heralding what Huang predicts to be the most successful launch in NVIDIA's storied history.