After a brief pause, additional info on the GeForce RTX 5090 has surfaced. The latest rumor alleges that the top Blackwell gaming GPU will feature a 512-bit memory bus pairing 16 GDDR7 dies to the graphics core, GB202. We’ve discussed this before and it was assumed that it was shelved in favor of a more conventional 384-bit design. However, that decision may have been overruled (the overruled was overruled).
According to “panzerlied” on the Chiphell forums, the GeForce RTX 5090 will leverage a dense memory configuration. The GB202 graphics die will (supposedly) be surrounded by 16x GDDR7 memory chips, up from 12x GDDR6X packages on the RTX 4090. Each GDDR6X die on the latter was connected to a 32-bit memory controller on the GPU.
A total of 16x GDDR7 memory dies implies a 512-bit configuration (32-bit x16) and a 32 GB memory buffer for the RTX 5090. These next-gen memory chips are expected to be clocked at 28 Gbps, up from 23 Gbps on the RTX 4080 Super. This results in a peak memory bandwidth of 1.79 TB/s, nearly 80% higher than the 1.01 TB/s sustained by the 4090.
According to @kopite7kimi, NVIDIA will use a triple-part PCB for the RTX 5090, consisting of the Main Board, IO Rigid Board, and the PCIe-slot component. A multi-part board would give the heatsink designers enough room to build a slimmer (and effective) cooling solution. After all, the RTX 4090 and 4080 have grown excessively thick.
NVIDIA has been testing unconventional PCB designs with its Founders Edition graphics cards for a while now. For example, the RTX 3090 and 4090 feature a PCB with a V-shaped lower end. This makes space for the second heatsink fan that blows in the opposite direction as its companion.
