NVIDIA’s upcoming flagship GPU, the GeForce RTX 5090, is set to feature a massive monolithic GB202 “Blackwell” die, according to reports from Kopite7kimi. This new GPU is expected to have up to 192 streaming multiprocessors (SMs), potentially housing up to 24,567 CUDA cores if it follows the design of previous AD102 “Ada” chips with 128 cores per SM.
Unlike NVIDIA’s HPC/AI chips that utilize a chiplet design, the GB202 “Blackwell” GPU is rumored to maintain a physically monolithic design for consumer-oriented GPUs. This design choice could result in a significant performance gap between the RTX 5090 and the smaller die RTX 5080, with the former being a powerhouse in terms of performance.
There have been speculations about the possibility of a chiplet-style design under the hood of the GB202 “Blackwell” GPU. This would allow NVIDIA to incorporate two GB203 dies in a monolithic package for improved inter-die communication without the complexities associated with traditional chiplet implementations. The GPU is also expected to be based on the TSMC N4P (5nm) process node, promising enhanced density and performance improvements.
In terms of features, reports suggest that the GeForce RTX 5090 could come with a 512-bit interface, along with a new cooling and PCB solution to support its performance. With rumors indicating that AMD may be stepping back from the ultra-high-end GPU segment, NVIDIA’s Blackwell GPUs are poised to solidify the company’s leadership in the gaming market.
The launch of the GeForce RTX 5090 is expected to follow closely after the RTX 5080, providing consumers with powerful options for their gaming needs. Overall, the GB202 “Blackwell” GPU is shaping up to be a significant advancement in NVIDIA’s GPU lineup, offering top-tier performance and features for gaming enthusiasts.
Article Source
https://wccftech.com/nvidia-geforce-rtx-5090-gpu-feature-massive-monolithic-gb202-blackwell-die/