Nvidia unveiled the Rubin CPX GPU earlier this week at the AI Technology Conference as part of its next-generation data center portfolio, which is designed to accelerate inference workloads. Rubin CPX, in particular, focuses on compute rather than bandwidth to form the other part of a "disaggregated" AI architecture that will kick off with Vera Rubin next year. Today, however, a closer look at its silicon suggests there might be more to this AI accelerator than Nvidia has let on, with some speculating that it could even lay the groundwork for an RTX 6090.
High Yield, an independent semiconductor analyst, recently extended his work to Rubin CPX, labelling the up-close die shot Nvidia released as part of the press material. This is just an artistic render—which could be inaccurate, so take the news with a healthy dose of skepticism— but that didn't stop High Yield from discovering something strange: the presence of graphics-specific hardware blocks. He found a full set of Raster Output Pipelines (ROP), raster units, and even display engines on Rubin CPX — all of which are components irrelevant to AI tasks.
Some quick assumptions about @Nvidia's new Rubin CPX. Quiet similar to GB202, with some obvious changes to the GPCs. Looks like it has full raster units (full GPU) with up to 256 ROPs. Could this be the RTX 6090? (even though Rubin was supposed to be AI only) pic.twitter.com/EBRYpH39VBSeptember 10, 2025
High Yield's labelling highlights 16 Graphics Processing Clusters (GPC), each with 6 Texture Processing Clusters (TPC), which would add up to 192 Streaming Multiprocessors (SM) — same as the Blackwell-based RTX 5090. But there are also 256 Raster Output Pipelines (ROP) on this die, along with four display pipes — again, taking up precious die area for an otherwise AI-only GPU. That makes Rubin CPX a different beast altogether because the GB202 powering the 5090 has just 170 ROPs and 12 GPCs, in comparison.
For a potential RTX 6090, Nvidia would almost certainly follow its usual practice of repurposing partially functional dies. Even with two GPCs disabled, Rubin CPX would still deliver around 28,672 CUDA cores and 224 ROPs, showing a notable increase over the RTX 5090’s 21,760 CUDA cores and 176 ROPs, respectively.
On the other hand, if Nvidia retains the 8 TPC-per-GPC layout carried over from Blackwell, the design could scale up to 32,768 CUDA cores. After accounting for typical yield-related cuts of roughly 10%, that still points to a gaming SKU in the 28,000+ CUDA core range, suggesting a 28–30% performance uplift gen-on-gen. That's even before factoring in clock speed or architectural refinements.
Moreover, the render also shows a 512-bit memory bus using GDDR7, and possibly 128 MB of L2 cache. That would push effective memory bandwidth to nearly 2 TB/s, comfortably ahead of the RTX 5090’s 1.8 TB/s (on a 512-bit interface as well). Not only that, but support for PCIe 6.0, and the presence of video engines further point toward a possibility that Rubin CPX—while an AI accelerator today—is also a stepping stone for what's to come tomorrow.
If Rubin CPX is indeed the precursor to the RTX 6090, it would represent one of the largest, most feature-complete dies Nvidia has ever shipped. With up to 256 ROPs, far more than the 176 ROPs in GB202, Rubin CPX's GPU could hypothetically deliver significant gains in traditional rasterization at higher resolutions, along with improvements in pixel throughput.
Rubin CPX will be released at the end of 2026, following Vera Rubin's launch. They will combine to form what Nvidia bills as its most advanced and efficient AI inference solution ever. The RTX 6090, on the other hand, should be announced at CES 2027 if Nvidia follows the patterns from previous years. Therefore, the timeline lines up, but it's far from confirmed. But right now we have to remain skeptical.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button!
