Nvidia (NASDAQ:NVDA) has announced new GPU architectures designed to support large-scale artificial intelligence operations. During its annual GTC event in San Jose, CEO Jensen Huang presented Blackwell Ultra and Rubin, two upcoming GPU models intended to enhance computational efficiency. These GPUs are expected to address increasing AI demands by offering high-speed processing capabilities. The company also introduced the Nvidia RTX PRO 6000 Blackwell Server Edition, emphasizing its role in enterprise applications. With these advancements, Nvidia continues to focus on AI-related hardware solutions while facing growing competition from other semiconductor firms.
When Nvidia introduced the original Blackwell architecture last year, the chips were described as significantly more powerful than their predecessors. However, early versions of the technology encountered overheating issues, which led to concerns among customers. In response, Nvidia has now introduced a liquid-cooled Grace Blackwell 200 NVL72 system, designed to enhance cooling efficiency and improve performance. Compared to previous iterations, the latest GPUs are advertised as having higher memory capacity and computational speed, particularly for AI model training.
What are the key features of Blackwell Ultra and Rubin?
Blackwell Ultra, set for release in late 2025, builds on the existing Blackwell architecture and integrates an upgraded interconnect technology called NVLink 72. This feature enables multiple GPUs to function as a unified processor, facilitating large-scale AI training and inference. The GPU architecture also includes high-bandwidth memory (HBM4E) with 288GB capacity, which is intended to improve efficiency in AI computations. Nvidia has also emphasized that the upcoming GPU will support enterprise applications requiring significant computing power.
Rubin, expected to launch in 2026, is named after astronomer Vera Rubin and is anticipated to surpass Blackwell Ultra in performance. The initial version will deliver 50 petaflops of AI processing power, while the Rubin Ultra variant will increase this capability to 100 petaflops. The company has positioned these GPUs as essential for handling large-scale generative AI models and industrial applications. Nvidia is also collaborating with TSMC to develop new packaging technologies, aiming to enhance future GPU designs.
How does Nvidia plan to expand beyond Rubin?
Following Rubin, Nvidia intends to introduce a new GPU architecture named Feynman, set for release in 2028. This development is expected to continue the company’s focus on AI-optimized hardware while addressing computational efficiency. Named after theoretical physicist Richard Feynman, this architecture will likely emphasize advancements in processing speeds and energy efficiency. Nvidia has not disclosed specific performance metrics for Feynman GPUs, but they are expected to integrate new memory technologies to accommodate increasingly complex AI models.
The company’s announcements coincide with heightened competition in the GPU market. Nvidia continues to dominate with a reported 80% market share, but rivals such as AMD (NASDAQ:AMD) are increasing efforts to develop competitive AI chips. Additionally, geopolitical factors, including semiconductor export restrictions, could influence Nvidia’s future growth. Despite these challenges, demand for AI-focused GPUs remains strong, driven by industries investing in machine learning and high-performance computing.
Nvidia’s latest GPUs reflect the growing need for computational power in AI development. Blackwell Ultra and Rubin aim to provide faster processing speeds and improved efficiency, addressing limitations seen in earlier models. With the addition of liquid cooling, Nvidia is addressing previous overheating concerns, while its partnership with TSMC suggests a long-term strategy to refine future architectures. As AI technology evolves, companies investing in large-scale computing will assess whether Nvidia’s GPUs meet their performance and cost requirements. The competition between Nvidia and other semiconductor firms will likely shape advancements in AI computing in the coming years.
