NVIDIA GeForce RTX 5090 To Feature Cut-Down GB202 GPU With 448-Bit Bus & 28 GB GDDR7 Memory
NVIDIA GeForce RTX 5090 graphics cards are rumored to have a cut-down 448-bit bus interface for up to 28 GB GDDR7 memory.
NVIDIA GeForce RTX 5090 Rumored To Get Cut-Down 448-Bit Bus Interface & 28 GB GDDR7 Memory
The new rumor comes from Chiphell Forum member, Panzerlied, who had previously stated the 512-bit bus interface and the new PCB design incorporating 16 GDDR7 memory modules. As per the latest information, the GB202 GPU featured on the NVIDIA GeForce RTX 5090 will come in a slightly cut-down configuration which would still mark an upgrade over the memory spec of the GeForce RTX 4090.
Related Story NVIDIA GeForce RTX 5090 Founders Edition GPU Gets Dual-Slot & Dual-Fan Cooling Solution
So starting with the details, the NVIDIA GeForce RTX 5090 graphics cards will feature a cut-down GB202 GPU with fewer cores than the fully enabled 192 SM chip. The GPU will also feature a cut-down memory configuration, featuring a 448-bit bus interface or 14 modules for up to 28 GB of GDDR7 memory capacities.
For comparison, the full version of the GB202 GPU will be equipped with a 512-bit interface and the new PCB will house up to 16 GDDR7 memory modules for up to capacities of 32 GB.
NVIDIA is also expected to utilize 28 Gbps GDDR7 memory modules from the get-go so if that is truly the case, we will see bandwidths of up to 1568 GB/s on the 448-bit configuration which is 56% more bandwidth than the RTX 4090 (384-bit @ 21 Gbps).
NVIDIA Blackwell 'GB202' GPU Specs 'Preliminary':
GPU NameGB202AD102 GPC12 (Per GPU)?12 (Per GPU) TPC8 (Per GPC)?6 (Per GPC) SM2 (Per TPC)?2 (Per TPC) Total SMs192?144 Sub-CoreTBD4 (Per SM) FP32128 (Per SM)?128 (Per SM) FP32+INT32TBD128 (Per SM) CUDA Cores24,567?18,432 WarpsTBD64 (Per SM) ThreadsTBD2048 (Per SM) L1 CacheTBD192 KB (Per SM) L2 CacheTBD96 MB (Per GPU) ROPsTBD32 (Per GPC) Memory StandardGDDR7GDDR6X Max Memory Bus512-bit384-bit Max Memory Cap32 GB24 GBA 512-bit interface would have given up to 1792 GB/s of bandwidth to the GPU which marks a