NVIDIA Blackwell: The New Frontier of AI Computing in Data Centers | Generative ai examples | Is chatgpt generative ai | | Turtles AI

NVIDIA has unveiled important news regarding its Blackwell platform, now up and running in data centers and ready to meet the growing demands of AI. Among the advances are the upgrade of NVLINK to 1.4 TB/s and the introduction of new system cooling technologies. The Blackwell GPU, equipped with 208 billion transistors and 8-site HBM3e memory, delivers 20 Peta FLOPS in FP4 AI, marking a breakthrough for low-precision AI computation. NVIDIA also shared the first image generated with FP4 computation, highlighting the potential of the new Quasar Quantization system.

Key Points:

1. Blackwell Operational: NVIDIA confirms Blackwell operation in data centers, ready for global expansion within the year.
2. Multi-GPU Integration: NVSwitch approach improves efficiency in communication between GPUs, reducing latency.
3. Energy Efficiency: New liquid cooling solutions promise significant reduction in operating costs.
4. Reduced Precision AI Generation: Image generated in FP4 demonstrates the ability to maintain high quality at higher speeds.

As the Hot Chips event approaches, NVIDIA has decided to dispel any doubts about the timing of the launch of the Blackwell platform, announcing that this innovative solution is already up and running in data centers and ready to be deployed globally by the end of the year. Speculation about possible delays is thus belied by the facts, as also demonstrated by the recent press session where NVIDIA presented Blackwell in action. Blackwell does not merely represent a single chip, but constitutes a complete platform that, as was the case with Hopper, includes a variety of designs designed for data center, cloud, and AI applications. The Blackwell product range includes not only GPUs, but also Grace CPUs, NVLINK switches, as well as several other key components such as BlueField-3, ConnettiX-7 and 8, Spectrum-4, and Quantum-3.

One of the core elements of Blackwell is the integration of GPUs with NVSwitch, which enables faster and more efficient communication between GPUs in a multi-GPU environment. In particular, the next-generation NVLINK switch offers double the bandwidth of its predecessors, reaching 1.8 TB/s and allowing up to 72 GPUs to be connected in a single GB200 NVL72 rack. This approach reduces latency and improves power efficiency, key aspects for inference in generative AI, where processing speed is crucial.

Another innovative aspect of Blackwell is liquid cooling, which will be a central theme of the NVIDIA presentations at Hot Chips. By using hot water directly on the chips, it achieves improved cooling efficiency and reduced operating costs by up to 28 percent, while also improving server life and opening up the possibility of reusing generated heat. This represents a significant advance over traditional cooling systems, which require a lot of energy to keep the liquid at a low temperature.

In the area of AI computing, NVIDIA presented a major innovation with the introduction of the Quasar Quantization system, which pushes the limits of reduced precision in AI computing. During the presentation, NVIDIA showed the first image generated with FP4 computation, a rabbit produced with Stable Diffusion. This image, although 4-bit, maintains surprisingly high quality compared to FP16 models while operating at significantly higher speeds. This innovation promises to make AI computation more efficient and faster, with potentially revolutionary applications in image generation and other AI models.

NVIDIA is not stopping there: the company is already using artificial intelligence to develop upcoming AI chips, optimizing Verilog code for the design and verification of processor circuits such as Blackwell. This approach allows NVIDIA to maintain a yearly cadence in the production of new GPUs, with anticipation of the Blackwell Ultra GPU in the coming year, which will feature 288 GB of HBM3e memory and even higher processing density.

NVIDIA’s technological advancement with Blackwell represents a key evolution in addressing the growing computing needs in artificial intelligence, with new solutions that will improve efficiency, speed, and sustainability in data centers.