Rubin CPX Die Shot Reveals Graphics Hardware Suggesting Future RTX 6090, AI-Ready with ROPs, Display Engines, and 512-bit Bus

Nvidia’s recently unveiled Rubin CPX GPU, presented at the AI Technology Conference, is positioned as a next-generation data center accelerator focused on AI inference workloads. Designed as part of a disaggregated AI architecture that will pair with the upcoming Vera Rubin system, Rubin CPX emphasizes raw compute power over memory bandwidth. However, a detailed analysis of the die shot released by Nvidia has sparked speculation that the chip may be more than just an AI-focused processor. Independent semiconductor analyst High Yield examined the artistic render of the Rubin CPX die and identified several graphics-specific components typically absent in AI accelerators. These include 256 Raster Output Pipelines (ROPs), four display engines, and a full set of Raster Processing Clusters (GPCs) and Texture Processing Clusters (TPCs). The layout suggests 16 GPCs with 6 TPCs each, totaling 192 Streaming Multiprocessors (SMs)—matching the configuration of the Blackwell-based RTX 5090. This is notable because AI GPUs generally omit such graphics-oriented hardware. The presence of 256 ROPs—significantly more than the 170 found on the GB202 GPU used in the RTX 5090—raises questions about the chip’s true purpose. ROPs are essential for traditional rasterization and pixel output, tasks irrelevant to AI inference. Yet, they consume valuable die space, suggesting that Rubin CPX may be engineered with future gaming applications in mind. If Nvidia repurposes partially functional dies from Rubin CPX for a consumer GPU, even with two GPCs disabled, the resulting chip could deliver around 28,672 CUDA cores and 224 ROPs—outpacing the RTX 5090’s 21,760 CUDA cores and 176 ROPs. With a full 8 TPC-per-GPC configuration, the design could scale to 32,768 CUDA cores. After accounting for typical yield losses of about 10%, a gaming variant could still offer over 28,000 CUDA cores—representing a 28 to 30 percent performance increase over the previous generation. The die also appears to feature a 512-bit GDDR7 memory interface and possibly 128 MB of L2 cache, potentially enabling memory bandwidth close to 2 TB/s—surpassing the RTX 5090’s 1.8 TB/s. Additional signs of a broader vision include support for PCIe 6.0 and integrated video engines. While Rubin CPX is officially intended for AI inference and will launch in late 2026, the architectural similarities to future consumer GPUs suggest it could serve as the foundation for the RTX 6090. If Nvidia follows its historical pattern, the RTX 6090 could debut at CES 2027. However, the die shot is an artistic rendering, and the actual silicon may differ. For now, the findings remain speculative, but they point to a possible convergence between Nvidia’s data center and consumer GPU strategies.