Zotac’s, no firstno secondBut a third accusation of accidentally listing a GPU too early reveals what Nvidia has been preparing behind the scenes for its RTX 50 GPU series. Zotac accidentally posted entries for the RTX 5090/D, RTX 5080, and RTX 5070 series on its website, which were quickly removed (H/T to video card z for discovery lists). While most of the aforementioned GPUs have already leaked, Zotac’s recently updated filter options are the first official source to confirm that the RTX 5090 will come with 32GB of GDDR7 memory.
GPU catalog includes RTX 5090 and its China-specific offerings RTX 5090D Variants. In the stack, we have the RTX 5080 and RTX 5070 series, which includes the base RTX 5070 and RTX 5070 Ti. It’s not like Nvidia launches Ti variants at the beginning of a new generation, as these usually come after updates. Still, this might be the GPU series we should look forward to Jen-Hsun Huang next month in International Consumer Electronics Show.
Leaked screenshots suggest that the RTX 5090 will be as large as its predecessor, taking up 3.5-4 slots of space in the case. In addition, Zotac has also updated the filtering standards of the graphics card part and added “GDDR7” and “32GB” options. Making some comparisons with what we know about the RTX 5090, it’s clear that Nvidia has equipped its flagship with 32GB of GDDR7 memory, compared to the RTX 4090’s 24GB.
To sum up everything we know so far, the RTX 5090 is rumored to be based on Nvidia’s GB202 chip – more than 744 square millimeters – Has 170 enabled SMs (192 total) and 32GB GDDR7 video memory. A slightly weakened RTX 5090D should also be built with GB202 and less streaming multi-processors – in 150 messages According to our calculations. Looking further down, the RTX 5080 using the GB203 chip drops to just 84 SMs (less than half of the RTX 5090) and 16GB of memory.
at last, RTX 5070 TitaniumAccording to Kopite reports, 70 SMs using GB203 chips will be provided. Otherwise, not much is known about the RTX 5070 series, except for some rumors that it will be available in February.
The RTX 50 series is expected to use TSMC’s 4NP process (5 nanometer level), which is said to be 30% denser than the 4N used on the RTX 40 series. It’s reasonable to assume that Nvidia may introduce FP4 and FP6 data types in Blackwell on the desktop, similar to its server versions (B100/B200). Still, there are a lot of details we don’t know, so let’s leave the architectural changes and pricing structure to Jensen next month.