The smoke of the chip war has never dissipated, and this time, NVIDIA has directly rewritten the rules of the game.“This is not a simple iteration, but a paradigm shift.” While we are still digesting the shock brought by the Blackwell architecture, Jensen Huang has already revealed his next trump card.
This chip, named Rubin CPX, marks a fundamental shift in NVIDIA’s product line.
It no longer pursues an “all-rounder” but is specifically optimized for two critical stages in large model inference: pre-fill and decode.
Paradigm ShiftToday’s large language model processing is divided into two distinctly different stages: pre-fill and decode.NVIDIA refers to these as the context stage and the generation stage.The traditional method is to use GPUs with massive HBM pools to run the entire process, but as GPU clusters scale from hundreds of thousands to millions, there is enough workload and scale to split these tasks into two more optimized architectures.
Rubin CPX is a co-processor specifically designed to handle these tasks.It boasts 30 PFLOPS of NVFP4 performance and 128GB of GDDR7 memory, with an exponential computing capability three times that of the GB300.Equipped with four NVENC/NVDEC engines, it directly targets video as a key workload.Industry LogicNVIDIA’s move reveals an important trend: AI computation is shifting from a training-centric to an inference-centric approach.As more large models enter the application phase, the demand for inference is exploding.Especially in long context windows and video content processing, traditional architectures are facing bottlenecks.
An hour of video content may require processing up to 1 million tokens, which completely breaks through the computational limits of traditional GPUs.
The innovation of Rubin CPX lies in its integration of video decoders, encoders, and long context inference processing into a single chip, providing unprecedented capabilities for long-format applications such as video search and high-quality video generation.Industry ChainEvery technological paradigm shift brings about a reshuffling of the industry chain.The emergence of Rubin CPX is not only significant for NVIDIA itself but also brings new opportunities to the entire AI ecosystem.In the data center field, chips specifically designed for inference tasks will change the design thinking of server architectures.Traditional homogeneous computing architectures will gradually shift towards heterogeneous architectures, posing new requirements for server manufacturers, data center operators, and cloud service providers.
At the application level, Rubin CPX will transform AI coding assistants from simple code generation tools into complex systems capable of understanding and optimizing large software projects.This is not just an improvement in efficiency but a qualitative change in capability.For the video industry, this means that AI processing of long video content will become possible.From video search to content generation, from personalized recommendations to automatic editing, the entire industry will usher in a new round of technological innovation.Rubin CPXWhile Rubin CPX brings exciting possibilities, we must also recognize that any technological paradigm shift takes time to complete.
NVIDIA has set the target listing time for the end of 2026, which means we still have more than a year to wait.
However, the popularization of heterogeneous computing architectures also faces challenges in software development, ecosystem building, and other aspects.
How to enable developers to better utilize this new type of computing resource, and how to build more efficient toolchains, are all issues that need to be addressed.From a broader perspective, AI computation is developing towards a more specialized and diversified direction.In the future, we may see more dedicated processors for specific scenarios and tasks, which will make the entire computing ecosystem richer and more complex.AIRedefinitionNVIDIA has once again proven its leading position in the field of AI computation.The release of Rubin CPX is not just the debut of a new product, but an important guide for the development direction of the entire industry.In this rapidly changing era, the only constant is change itself.The specialization and refinement represented by Rubin CPX are likely to become an important trend in the future development of AI computation.How to seize the dividends brought by technological changes, and how to occupy a favorable position in the new round of competition, are all questions worth our deep consideration.The future of AI is being redefined, and this time, NVIDIA is once again at the forefront.NVIDIA suddenly turns to silicon carbide, and the industry chain concept collectively erupts!NVIDIA, major announcement! Jensen Huang’s latest statementJensen Huang: Robots are NVIDIA’s next main battlefieldNote: This article is for industry personnel reference only and does not constitute any investment advice. Images and data are sourced from public information. If there is any infringement, please contact for deletion.