Zotac’s, not first, not second, but third count of accidentally listing GPUs prematurely has shed light on what Nvidia has been preparing behind the scenes for its RTX 50 lineup of GPUs. Zotac had accidentally put up entries for the RTX 5090/D, RTX 5080, and the RTX 5070 family across its website, which were promptly taken down (H/T to Videocardz for spotting the listings). While most of the aforementioned GPUs have already been leaked, Zotac’s recently updated filtering options are the first official source to confirm that the RTX 5090 will ship with 32GB of GDDR7 memory.
The catalog of GPUs includes the RTX 5090 in tandem with its China-specific RTX 5090D variant. Down the stack, we have the RTX 5080 and the RTX 5070 family featuring the base RTX 5070 and the RTX 5070 Ti. It isn’t like Nvidia to launch the Ti variants at the beginning of a new generation, as these products typically arrive later as a refresh. Still, this is probably the lineup of GPUs that we should expect from Jensen Huang next month at CES.
The leaked screenshots hint that the RTX 5090 will be as massive as its predecessor, taking up 3.5-4 slots of space in your case. In addition, Zotac updated the filtering criteria in its Graphics Card section to include “GDDR7” and “32GB” options. Drawing a few parallels with everything we know about the RTX 5090, it should be clear that Nvidia is packing its flagship with 32GB of GDDR7 memory – up from 24GB on the RTX 4090.
Summing up everything we know thus far, the RTX 5090 is rumored to be based on Nvidia’s GB202 chip – at over 744mm-squared – with 170-enabled SMs (out of 192 in total) and 32GB of GDDR7 memory. The slightly nerfed RTX 5090D should also be built using GB202 with fewer Streaming Multiprocessors – at 150 SMs per our calculations. Further down the list, the RTX 5080 using GB203 silicon drops to just 84 SMs – which is less than half that of the RTX 5090 – and 16GB of memory.
Lastly, the RTX 5070 Ti, per Kopite, will offer 70 SMs featuring the GB203 chip. Other than that, not much is known about the RTX 5070 family apart from a few rumors that slate it for retail in February.
The RTX 50 lineup is expected to employ TSMC’s 4NP process (5nm-grade) – said to deliver 30% higher density than 4N used on the RTX 40 series. It is fair to assume Nvidia might introduce the FP4 and FP6 data types natively with the Blackwell on desktop, similar to its server counterpart (B100/B200). Still, there are a lot of nitty-gritty details we don’t know so let’s leave the architectural changes and pricing structure for Jensen next month.