Lang Chai King
March 20, 2025 03:29
NVIDIA introduces Blackwell Ultra, a platform designed for the era of AI reasoning, to provide performance improvement for education, training and test time scaling.
NVIDIA has announced the launch of Blackwell Ultra, a new acceleration computing platform adjusted to meet the evolution of AI reasoning. According to NVIDIA, this platform is designed to improve the function of the AI system by optimizing training, training and test time scaling.
AI scaling development
Over the past five years, the requirements for pre -adjustment of AI have surged by 50 million people, leading to considerable development. However, the focus is now switching to a refined model to improve reasoning ability. This includes scaling after training, which uses domain stars and synthetic data to improve understanding of AI’s conversation skills and subtle situations.
A new scaling method called ‘test time scaling’ or ‘long -term accident’ has emerged. This approach dynamically increases computing resources during the AI reasoning, enabling deeper reasoning. Unlike the existing models that generate response in a single pass, these high -end models can move closer to autonomous intelligence and think and improve their answers in real time.
Blackwell Ultra Platform
The Blackwell Ultra platform is the core of NVIDIA’s GB300 NVL72 system and consists of a liquid cooling rack scale solution connecting the 36 NVIDIA Grace CPU and the 72 Blackwell Ultra GPU. This setting greatly improves AI reasoning performance by forming a large GPU domain with a total NVLINK bandwidth of 130TB/s.
Blackwell Ultra supports large AI models and complex tasks with up to 288GB of HBM3E memory per GPU to reduce performance and wait time. The Tensor Core provides 1.5 times more AI computing flop than the previous model to optimize memory usage and enable breakthrough innovation in AI research and real -time analysis.
Inference and networking improvement
NVIDIA’s Blackwell Ultra features a PCIe Gen6 connection with NVIDIA ConnectX-8 800g Supern and improves network bandwidth to 800GB/s. This increased bandwidth improves the performance of NVIDIA DYNAMO, an open source library that expands AI services and efficiently manages GPU node workloads.
The separated serving of Dynamo separates the context and production steps of the large language model (LLM) reasoning to optimize the performance to reduce costs and improve the scalability. The GB300 NVL72 is completely integrated with NVIDIA’s Quantum-X800 and Spectrum-X platform with a total data throughput of 800GB/s per GPU to meet the needs of modern AI plants.
Effect on the AI factory
The introduction of Blackwell Ultra is expected to greatly improve the output of the AI factory. The NVIDIA GB300 NVL72 system has increased 10 times per user and 5 times the throughput per mega watt to 50 times the overall AI factory output performance.
This development of AI reasoning will promote real -time insights, strengthen prediction analysis, and improve AI agents in various industries, including finance, medical and e -commerce. Tissue can handle larger models and workloads without damaging speed, allowing advanced AI functions to be more practical and accessible.
NVIDIA Blackwell Ultra Products is expected to be provided by partners in the second half of 2025 and is supported by major cloud service providers and server manufacturers.
Image Source: Shutter Stock