summary
Nvidia is planning a new class of GPUs called Rubin CPX, designed specifically for the computational analytic phase of AI models. The strategy known as split inference is supported by Nvidia's Blackwell Ultra Architecture's new benchmark record, which uses a similar approach in software.
When AI models tackle complex tasks, they usually work in two phases. For example, if you ask your language model to summarise a long book, you must first read and analyze the entire text. This is a computationally intensive “analysis” or “context phase.” Only then will the “generating phase” begin, and the summary is created for each word. Nvidia claims that these two phases have different hardware requirements and that Rubin CPX is a new chip built specifically for its demanding analytics phase.
According to the official announcement, the Rubin CPX is a derivative of the Rubin product line planned for 2025 and is expected to be released at the end of 2026. It can be used as a plug-in card or as a standalone computer for data centers.
Professional chips for huge context windows
Nvidia says Rubin CPX shines for tasks that require you to process huge amounts of data at once. This includes AI applications that require millions of tokens as context, such as analyzing the entire codebase and generating videos. CEO Jensen Huang calls the first CUDA GPU that has built CPX specifically for “Massive-ContextAI”.
advertisement
The technical foundation for this is Nvidia's “decomposed reasoning” strategy. As detailed in the Technical Blog post, the Computation Bound Context Phase places requests in hardware differently from the next generation phase of memory bandwidth. Rubin CPX uses a monolithic die design, supplies 30 NVFP4 computing, packs 128 GB of GDDR7 memory, providing three times more attention layer acceleration compared to Blackwell.
Blackwell backs up the approach
Nvidia points to recent benchmark results to support its strategy. In the latest round of MLPERF Inference v5.1 industry standard, Nvidia submitted the results for the Blackwell Ultra for the first time and set up a new record.
According to NVIDIA, the Blackwell Ultra, running on a GB300 NVL72 rack system, achieved up to 45% higher per GPU performance than the standard Blackwell. The throughput of the new DeepSeek-R1 benchmark is reportedly five times higher than the previous hopper architecture.
Nvidia points out that these tests already use a software approach to splitting inference. For the LLAMA 3.1 405B model, “decomposed serving” was used to split the context and generation phases across different GPUs. The method, managed by the Nvidia Dynamo Framework, is nearly 1.5 times boosted per GPU compared to the traditional method. With Rubin CPX, this software approach gives you a dedicated hardware solution. Nvidia says partners like Cursor (Code Editor), Runway (Video AI), and Magic (AI Agent) have already evaluated the technology for their own use cases.
