Last week, it was disclosed that the NVIDIA RTX 5090 PCB will be more complicated than that of the RTX 40 series. Unlike the previous models, the RTX 50 series will employ three interconnected boards instead of a single one. This design is exclusive to NVIDIA’s Founders Edition model and might not be adopted by reference models from other manufacturers.
NVIDIA likely had compelling reasons for choosing such a complex cooler design for the RTX 5090. Prototypes of the GB202 Blackwell GPU, apparently intended for the forthcoming flagship RTX 5090, indicate it will occupy only two slots. This marks a notable decrease from the bulky RTX 4090 Founders Edition or custom 4090 models, which generally take up 3 to 4 slots.
Speculating about the RTX 5090’s specifications might be premature, but here’s what we know so far: The full GB202 GPU is anticipated to have up to 192 Streaming Multiprocessors. Like the RTX 4090, the 5090 will likely have fewer cores than the full chip, though the exact number remains unknown.
More certain is that the RTX 5090 will likely use a 512-bit memory bus. The memory capacity will depend on whether NVIDIA opts for 2GB or 3GB modules, according to JEDEC specifications. This means the card could feature either 32GB or 24GB of memory. Rumors suggest the PCB is designed to support 16 GDDR7 modules, making the 32GB option more likely.
NVIDIA is not expected to make any announcements during the pre-Computex keynote this weekend. However, more details are likely to emerge over the summer, as the RTX 50 series is expected to be available by the end of the year.
Source: Kopite7kimi