.png "blank Coverage")
Nvidia, propelled by its groundbreaking H100 AI chip, has cemented its position as a multitrillion-dollar company, potentially surpassing tech giants like Alphabet and Amazon. Now, Nvidia is poised to further extend its lead in the industry with the introduction of the new Blackwell B200 GPU and GB200 "superchip."
According to Nvidia, the B200 GPU boasts up to 20 petaflops of FP4 horsepower from its 208 billion transistors. Additionally, the GB200, which combines two B200 GPUs with a single Grace CPU, promises 30 times the performance for LLM inference workloads while also offering significant efficiency gains, reducing cost and energy consumption by up to 25 times compared to the H100.
Training a massive 1.8 trillion parameter model, which previously required 8,000 Hopper GPUs and 15 megawatts of power, can now be accomplished with just 2,000 Blackwell GPUs consuming only four megawatts. In benchmark tests with a GPT-3 LLM model, the GB200 demonstrates seven times the performance of an H100 and four times the training speed.
Key improvements in the Blackwell architecture include a second-gen transformer engine that doubles compute, bandwidth, and model size by utilizing four bits for each neuron. Additionally, a next-gen NVLink switch enables 576 GPUs to communicate with each other, boasting 1.8 terabytes per second of bidirectional bandwidth.
To address communication bottlenecks in large GPU clusters, Nvidia developed a new network switch chip with onboard compute, providing 3.6 teraflops of FP8. The Blackwell architecture introduces both FP4 and FP6 capabilities.
Nvidia envisions companies deploying large quantities of these GPUs, packaging them in larger designs such as the GB200 NVL72 rack, which integrates 36 CPUs and 72 GPUs for a total of 720 petaflops of AI training performance or 1,440 petaflops of inference. Each rack supports a 27-trillion parameter model, making it ideal for advanced AI applications.
Cloud service providers including Amazon, Google, Microsoft, and Oracle are already planning to offer the NVL72 racks in their offerings. Nvidia also offers comprehensive solutions like the DGX Superpod for DGX GB200, combining eight systems into one for massive computing power.
With the ability to scale to tens of thousands of GB200 superchips, Nvidia's systems are equipped with advanced networking technology, enabling seamless communication and collaboration in large-scale AI deployments.
While gaming GPU announcements are not expected during Nvidia's GPU Technology Conference, the Blackwell GPU architecture will likely power future RTX 50-series desktop graphics cards, further solidifying Nvidia's dominance across multiple industries.
 |
| Training a massive 1.8 trillion parameter model, which previously required 8,000 Hopper GPUs and 15 megawatts of power, can now be accomplished with just 2,000 Blackwell GPUs |