Nvidia's next-generation GPUs are on the horizon, and enthusiasts are eagerly awaiting the latest developments. In this article, we delve into the recent revelations made by the renowned Nvidia insider, Kopite7kimi, regarding two high-end GPUs: the GB100 and GB202. These GPUs promise significant upgrades, including increased SM counts and L2 cache, among other improvements.
Upcoming Nvidia GPUs
As we look forward to the year 2025, Nvidia's next-generation graphics cards are generating considerable buzz. Although they are not expected to hit store shelves until early 2025, details about some high-end SKUs have begun to surface. Kopite7kimi, a well-known Nvidia leaker, has provided valuable insights into two flagship GPUs that are set to redefine gaming and datacenter performance.
Revamped Naming Scheme
Nvidia is not just upgrading its GPUs; it's also revamping its naming scheme. Instead of sticking with the GB1xx nomenclature for all SKUs, the company is adopting a more nuanced approach. It will use GB1xx for top-tier data center GPUs and GB2xx for gaming and other variants, signifying a clear distinction between the two product categories.
GB100: Gaming and Datacenter Powerhouse
The GB100 is poised to be a gaming and datacenter powerhouse. It boasts 10 Texture Processing Clusters (TPCs) in each of its 8 Streaming Multiprocessors (SMs), resulting in a total of 160 SMs (assuming 2 SMs per TPC). This configuration promises exceptional performance and rendering capabilities. Additionally, the GB100 features a 512-bit memory bus, a critical component for memory access and bandwidth. Previous rumors suggest that it will employ a Multi-Chip Module (MCM) design, similar to what AMD implemented with RDNA 3.
GB102: Elevating the Gaming Experience
The GB102 takes gaming to new heights with its 12 SMs, each equipped with 8 TPCs, totaling an impressive 192 SMs. This represents a remarkable ~33% increase in SM count compared to the AD102 (144). While the exact number of CUDA cores remains uncertain at this stage, it could potentially surpass the Ada Lovelace architecture's 128 CUDA cores per SM. The type of memory that GB102 will utilize is yet to be specified, but GDDR7 is a possibility, with Nvidia likely reserving it for its datacenter SKUs.
GeForce RTX 5090: The GB202 GPU
One of the highly anticipated products, the GeForce RTX 5090 (tentative name), will be powered by the GB202 GPU. It is expected to ship with a reduced number of CUDA cores and a narrower memory bus (384-bit) when compared to past cards as a reference point. Despite these changes, early rumors suggest that it could be up to 2.6 times faster than its Ada Lovelace counterparts. However, this performance boost will also depend on Nvidia's choice of the manufacturing node for Blackwell.
The Manufacturing Process
The choice of the manufacturing process is crucial for Nvidia's upcoming GPUs. A custom version of TSMC's N3 manufacturing process seems likely if there is sufficient capacity available. Alternatively, Nvidia might have to stick with the TSMC 4N process for its GeForce cards, while opting for an N3 variant for its data center products. This strategic decision could play a pivotal role in the performance and efficiency of these GPUs, especially in the absence of a high-end RDNA 4 equivalent from the competition.
Conclusion
Nvidia's GB100 and GB202 GPUs represent a significant leap forward in gaming and datacenter technology. With increased SM counts, revamped naming schemes, and the potential for cutting-edge manufacturing processes, these GPUs promise to deliver an extraordinary computing experience. While the official release is still on the horizon, these revelations from Kopite7kimi provide a tantalizing glimpse into Nvidia's future and leave enthusiasts eagerly awaiting more details.
