Micron has updated its GPU memory roadmap, adding high-bandwidth GDDR7, HBM4, and HBM4E solutions for next-gen graphics cards. The HBM side will remain limited to the data center. However, GDDR7 will likely be leveraged by NVIDIA in its RTX 5090 and other high-end 50 series gaming GPUs. Lower-end RTX 5060 and 5070 should continue to rely on fine-tuned GDDR6X solutions.
Micron plans two iterations of GDDR7 memory for gaming GPUs. The first will debut in the second half of 2024, just ahead of the 12-Hi HBM3E data center memory. It will be a major uplift over existing GDDR6X solutions, offering an effective clock of 32 Gbps versus 21 Gbps and 22.5 Gbps on the RTX 4090 and 4080, respectively.
Assuming that the GeForce RTX 5090 features a 384-bit bus, the use of 2,000 MHz GDDR7 (32 Gbps effective) memory would result in an external memory bandwidth of 1,536 GB/s, 52% higher than the 1,008 GB/s managed by the RTX 4090. In terms of density, GDDR7 will offer 24Gb memory dies, up from 16Gb with GDDR6X.
Both these advances should technically allow the GeForce RTX 5090 to feature up to 36GB of GDDR7 memory clocked at 2,000MHz or 32Gbps (effective). The resulting bandwidth would be 1,536GB/s, much higher than any GDDR6/GDDR6X solutions currently on the market.
Speaking of GDDR6X, NVIDIA may decide to equip the RTX 40 Supers with faster 24Gbps memory kits. That would offer a healthy bandwidth (and performance) upgrade without increasing the BOM by much. The RTX 4080 Super (and perhaps the 4070 Ti Super) could get this upgrade. Unlikely for the RTX 4070 Super as it targets 1080p and 1440p gaming.