According to a recent report from Nvidia leaker kopite7kimi, the upcoming GB202 GPU, which is likely to be used on the RTX 50 series flagship card, could offer significantly improved performance compared to the AD102 GPU. The GB100 GPU, designed for data center use, is expected to feature 8 graphics processing clusters (GPCs), each consisting of 10 texture processing clusters (TPCs). With each TPC including 2 streaming multiprocessors (SMs), the GB100 GPU would have a total of 160 SMs.
In comparison, the previous GH100 GPU had 144 SMs, indicating an increase in core count. The GB202 GPU, rumored for gaming and workstation cards, is said to have 12 GPCs with 8 TPCs each, resulting in a total of 192 SMs. This configuration would provide a significant boost in performance, featuring 24,576 CUDA cores, surpassing the current AD102 GPU used in the RTX 4090 and RTX 6000 ADA.
While Nvidia has yet to fully utilize the 144 SM count, the flagship workstation graphics card currently offers 142 SMs enabled. Unfortunately, details regarding the configuration of Tensor and RT Cores in the GB202 GPU are still unknown.
Kopite7kimi also supported previous rumors suggesting that the GB202 GPU would have a 512-bit memory bus. However, it's important to note that this memory arrangement may not be used in the RTX 5090. Furthermore, there are conflicting rumors about the type of memory, with some suggesting GDDR7 and others mentioning GDDR6X/W.
If these rumored specifications hold true, it appears that we are in for another significant leap in GPU performance. However, it's essential to approach these details with caution, as they do not provide the complete picture. Further information is needed to fully understand the capabilities and features of the upcoming GB202 GPU.
Title: The Upcoming Nvidia GB202 GPU Could Surpass the AD102 in Performance
Source: Nvidia Leaker kopite7kimi
– GPU: Graphics Processing Unit, a specialized electronic circuit that accelerates the creation and rendering of images, video, and animations.
– GPC: Graphics Processing Cluster, a subdivision within a GPU that contains several streaming multiprocessors.
– TPC: Texture Processing Cluster, a subdivision within a GPC that manages the processing of textures in a GPU.
– SM: Streaming Multiprocessor, a unit within a GPU responsible for executing instructions and performing computations.
– CUDA Cores: Parallel processing units within Nvidia GPUs designed to handle specific tasks.
– Tensor Cores: Specialized hardware units in Nvidia GPUs that accelerate deep learning algorithms.
– RT Cores: Ray Tracing Cores, dedicated hardware units in Nvidia GPUs that accelerate real-time ray tracing calculations.
– GDDR: Graphics Double Data Rate, a type of memory commonly used in GPUs for high-bandwidth data transfer.
– KitGuru: A technology news website covering hardware, gaming, and consumer electronics.
– kopite7kimi (@kopite7kimi) September 28, 2023