During this season, numerous rumors are circulating. Kopite7kimi, a highly reliable NVIDIA hardware leaker, has begun revealing new information, introducing a subtle twist to the unfolding narrative. Returning to the rumors about the 512-bit memory bus interface, Kopite7kimi initially asserted that the flagship gaming GPU in the upcoming series would have a 512-bit memory bus, a claim supposedly corroborated by various sources. However, a few months later, the narrative shifted, suggesting a 384-bit interface instead. Last week, the same leaker hinted at the possibility of RTX 50 and 40 series sharing similar memory bus widths, indicating minimal changes for the flagship GPU, GB202.
Presently, Kopite7kimi is reintroducing the possibility of a 512-bit configuration without providing specific details. It’s uncertain if NVIDIA is considering varied memory specifications or if the leaker is speculating based on knowledge. Considering the historical accuracy, one might incline towards the former possibility. However, the details remain somewhat elusive at this point.
Crucially, Kopite now asserts that the upcoming series will adhere to GDDR7 28 Gbps memory, surpassing GDDR6X specifications significantly. This positions NVIDIA with ample room for future enhancements, considering that GDDR7 memory is usually marketed with speeds of 32+Gbps.
The inclusion of 3GB GDDR7 modules is unlikely in the initial RTX 50 lineup, but they are in development and expected to launch next year, coinciding with the potential release of NVIDIA’s mid-range and entry-level GPUs. This hints at the possibility of new memory capacity choices, which shouldn’t be dismissed.
In the worst-case scenario, NVIDIA will maintain the same memory bus widths for their next-gen gaming GPUs, resulting in significantly increased bandwidth due to the faster GDDR7 memory. However, should NVIDIA, for some reason, consider the 512-bit option, it would mean the GPU supporting even greater memory capacity and offering nearly 80% higher bandwidth than AD102.
In conclusion, the leaker hinted that the GB203 GPU could be half the size of GB202, remaining ambiguous about whether GB202 employs a multi-chip package. The current flagship AD102 exhibits an 80% increase in cores compared to AD103, elevating RTX 4090(D) GPUs to a markedly higher performance tier. If NVIDIA follows a comparable strategy with Blackwell, we could potentially observe a substantial performance disparity once again between the RTX 5090 and 5080 series, particularly given the ongoing speculations about a 512-bit bus.
28Gbps.
— kopite7kimi (@kopite7kimi) March 11, 2024