New Blackwell Ultra GPU
Nvidia's Blackwell Ultra chip is a major advancement in artificial intelligence, offering 288 GB of memory per chip. This architecture is designed to meet the growing needs of advanced AI applications, such as agentic AI and reasoning AI. With 20 petaflops of AI performance, it enables models to solve complex problems autonomously while reducing operational costs through better memory management and increased efficiency.
Vera Rubin and Rubin Ultra GPUs
Nvidia has unveiled its future roadmap with the Vera Rubin and Rubin Ultra architectures. Vera Rubin, planned for 2026, promises 50 petaflops of performance in FP4, while Rubin Ultra, expected in 2027, will double this performance with 100 petaflops and 1 TB of memory per GPU. Rubin Ultra will use four interconnected GPUs, delivering a major technological leap with significantly improved inference and training performance. These architectures aim to meet the growing needs of advanced AI applications and complex simulations.
NVIDIA Dynamo
Nvidia has launched Dynamo, an open-source operating system for AI, designed to accelerate and optimize AI reasoning models in AI factories. Dynamo enables efficient management of inference requests across large fleets of GPUs, reducing operational costs while increasing performance. Through decomposed processing and intelligent routing techniques, Dynamo simplifies the implementation and management of AI infrastructures, facilitating the adoption of AI in various sectors.
Advances in Robotics and Automotive
Nvidia has teased significant advances in robotics and automotive, with autonomous systems and self-driving technologies. The partnership with General Motors to integrate AI into vehicle design and production is a notable example. These innovations could transform the automotive industry by integrating AI into production, design, and the driving experience. Autonomous systems promise increased safety and improved efficiency in factories and on the road.
Quantum Computing
Nvidia hosted a “quantum day” at GTC to explore real-world applications of quantum computing, potentially revolutionizing fields like cryptography and molecular simulation. The goal is to link Nvidia GPUs with quantum advances to solve problems that are intractable for classical computers. This initiative aims to open new avenues for scientific research and technological innovation, harnessing the unique capabilities of quantum computing to tackle complex challenges.
The Arrival of Personal AI Supercomputers (DGX Spark and DGX Station)
Nvidia introduced the DGX Spark and DGX Station personal AI supercomputers, designed to enable users to prototype and deploy complex AI models locally. DGX Station uses the Blackwell Ultra chip, delivering 20 petaflops of AI performance, while DGX Spark is powered by the GB10 chip and offers considerable power in a compact form factor. These systems aim to democratize access to AI for researchers and developers, allowing them to work on advanced AI projects without relying on expensive cloud infrastructure. Partnerships with several manufacturers such as ASUS have also been announced.
A new range of AI servers (GB300 NVL72)
Nvidia's GB300 NVL72 system combines 72 Blackwell Ultra GPUs and 36 Grace CPUs, delivering 50 petaflops of inference power. Expected to be available in the second half of 2025, this system is designed to meet the growing needs of data centers and large-scale AI applications. It offers compatibility with existing infrastructures while improving performance and memory capacity compared to previous models, facilitating integration into current production environments.
0 Comments