Nvidia GeForce RTX 5000 graphics cards could launch with GDDR7 memory and the same bus width as the RTX 4090

gddr7Prolific leaker Kopite7kimi asserted on multiple occasions that the GeForce RTX 5090's GB202 GPU would have 512-bit memory bandwidth. It was based on the assumption that the graphics card would use 24 Gbps GDDR6x memory. However, new information from RedGamingTech says it will use 32 Gbps GDDR7 memory on a 384-bit bus for a total memory bandwidth of 1.5 TB/s, assuming it uses 24 GB of VRAM.

RedGamingTech adds that the amount of L2 cache could be upped to 96 MB. Rumour has it that it could be as high as 128 MB, but that is likely reserved for AI-centric products. Performance uplift predictions are all over the place, with figures ranging between 2.6x and 2.0x over the RTX 4090 (US$2,099 on Amazon). However, RedGamingTech says that could apply only to raytracing.

Although not stated explicitly, the GeForce RTX 5000 series will stick to Nvidia's tried-and-tested monolithic die. HBM3-based chiplet tech will be reserved for AI and HPC products. It will be manufactured on TSMC's 3 nm process node. It is unknown if it is an off-the-shelf node (N3E/N3P) or a customized version (3N). It is also speculated that only some high-end products will be based on 3 nm, with the remaining sticking to TSMC 4N.

A leaked roadmap from earlier talked about Nvidia's plans to launch the RTX 50 Blackwell series in 2025, which was subsequently revised to H2, 2024 by another report. The latter sits in line with Nvidia's usual launch schedule, but the former will allow Nvidia to fine-tune its hardware better. With no high-end RDNA4 graphics cards due to be launched anytime soon, Nvidia has a lot of breathing room.