Nvidia Vera Rubin, Rubin Ultra confirmed for 2026 and 2027. Up to 3.3x and 14x Blackwell Ultra GB300 | Types of computer hardware | Gpu vs cpu laptop | 4 main parts of a computer | Turtles AI

NVIDIA is preparing to launch its new AI systems based on the Rubin and Rubin Ultra platforms, featuring Vera GPU and CPU chips, with extraordinary performance, HBM4e memory, and a revolutionary NVLink interconnect, starting in 2026.

Key Points:

• The new Rubin and Rubin Ultra platforms will arrive in 2026 and 2027, respectively, bringing huge improvements in compute and memory performance.
• The NVL144 and NVL576 systems will support inference performance of up to 15 ExaFLOPS, with memory capacity of up to 1 TB of HBM4e.
• The new Rubin GPUs will feature lattice chips and will be joined by Vera processors, with ARM architecture and advanced interconnects.
• The new NVLink 7 interface and CX9 will deliver unprecedented bandwidth, improving overall efficiency in data centers.

NVIDIA has unveiled ambitious plans for the future of its AI platforms, marking a major step in the evolution of computing infrastructure for advanced data centers. The Rubin and Rubin Ultra systems, powered by new Rubin GPU chips and the Vera CPU, are designed to meet the growing demands of AI and deep learning applications, delivering unprecedented performance and memory capacity. The first of the two systems, the Vera Rubin NVL144, will be available in the second half of 2026, followed by the Rubin Ultra NVL576 in the second half of 2027.

The Rubin NVL144 system will feature a dual-GPU lattice chip configuration, capable of up to 50 PFLOPS of FP4 performance and featuring 288 GB of HBM4e memory, a significant improvement over the previous-generation Blackwell. These chips will be joined by the new Vera CPU, based on a custom ARM architecture with 88 cores and 176 threads. Communication between the components will be via an NVLink-C2C interconnect, with a bandwidth of up to 1.8 TB/s, ensuring high computing performance for real-time AI applications.

The Rubin NVL144 will not only improve FP4 computing capacity, with a 3.3x increase compared to the previous generation, but will also achieve a significant increase in FP8 performance, with 1.2 Exaflops dedicated to training. The new HBM4 memory, with a bandwidth of 13 TB/s, stands out for its speed and ability to handle the heaviest workloads. The new NVLink architecture and CX9 support will ensure an even faster connection, doubling the throughput capacity compared to the previous generation, reaching 260 TB/s.

In 2027, the Rubin Ultra NVL576 will be an enhanced version of the Rubin NVL144 system, with twice the GPU chips and 1TB of total HBM4e memory. With four chips per GPU unit, Rubin Ultra will have 100 PFLOPS of FP4 compute and 5 Exaflops of FP8 training. The NVL576 system will leverage a new NVLink7 interface, which will be six times faster than the one used in the Rubin NVL144, and will have a total throughput of 1.5 PB/s. The system’s memory capacity will grow exponentially, reaching 365TB of “fast” memory, a significant increase from the 75TB of the NVL144.

The Rubin Ultra platform will be designed to meet the compute and memory requirements of the most demanding AI solutions, such as those used in large data centers. With support for 576 GPUs in a single unit, it will be possible to run inference and training operations at an unprecedented speed, with performance reaching a 14x increase compared to the previous generation. This improvement is also reflected in the CX9 architecture, which will double the data transfer capacity between units, reaching 115.2 TB/s.

These technological advances are part of a context of continuous innovation by NVIDIA, which, with the subsequent transition to Feynman architectures, aims to further refine its solutions for artificial intelligence and deep learning workloads. NVIDIA’s roadmap is shaping up to be a real challenge for the industry, with platforms that will redefine the performance standards for AI technologies and next-generation data centers.

NVIDIA, therefore, is preparing to radically transform the computing infrastructure for AI, with a vision that goes well beyond the current limits, pushing the frontiers of innovation.