Story perspectives

NVIDIA Unveils Rubin CPX GPU for AI Revolution

9/10/2025

Summary

Story summary

NVIDIA launched the Rubin CPX GPU for large-scale AI inference, processing over one million tokens for tasks like software coding and video generation.
It offers 30 petaFLOPs of NVFP4 compute power and 128 GB of GDDR7 memory, enhancing performance during the context phase of inference.
The Rubin CPX integrates into the Vera Rubin NVL144 CPX platform, combining 144 GPUs and 36 CPUs for 8 exaFLOPs of AI compute.
Its architecture improves efficiency by separating context and generation phases, reducing latency.
NVIDIA anticipates $5 billion returns for every $100 million invested in Rubin CPX infrastructure, with a release planned for late 2026.

History

Historical Context

In 2012, NVIDIA launched the Kepler architecture, achieving 4.5 teraflops of performance and significantly advancing GPU computing. This innovation facilitated AI and deep learning applications, with major companies like Google and Facebook utilizing GPUs for data processing. The introduction of the Rubin CPX GPU, promising 30 petaFLOPs, highlights the ongoing evolution in computational power to meet the increasing demands of AI technologies.

1 of 2

Story summary

NVIDIA launched the Rubin CPX GPU for large-scale AI inference, processing over one million tokens for tasks like software coding and video generation.
It offers 30 petaFLOPs of NVFP4 compute power and 128 GB of GDDR7 memory, enhancing performance during the context phase of inference.
The Rubin CPX integrates into the Vera Rubin NVL144 CPX platform, combining 144 GPUs and 36 CPUs for 8 exaFLOPs of AI compute.
Its architecture improves efficiency by separating context and generation phases, reducing latency.
NVIDIA anticipates $5 billion returns for every $100 million invested in Rubin CPX infrastructure, with a release planned for late 2026.

1 / 2