The much-anticipated details of NVIDIA's GeForce "Blackwell" series GPU specifications have been leaked by a trusted insider, Kopite7kimi. According to the leaked information, NVIDIA is gearing up to launch a new lineup of GeForce RTX 50 series graphics cards, which will be based on these newly leaked chips. The series is set to include five different chips, with the GB202 being the largest and most powerful, followed by the GB203, GB205, GB206, and the GB207.

The flagship GB202 is expected to be the powerhouse behind the potential RTX 5090, NVIDIA's successor to the RTX 4090. This chip will come with 192 streaming multiprocessors (SM) and 96 texture processing clusters (TPCs) spread across 12 graphics processing clusters (GPCs). Assuming the same 256 CUDA cores per TPC as seen in previous NVIDIA gaming GPUs, the GB202 will have a total of 24,576 CUDA cores. Additionally, the GPU will implement next-generation GDDR7 memory with a massive 512-bit memory bus, offering a memory bandwidth of 1,792 GB/s.

The GB203, as the next in line, is poised to be the successor to the current AD103, featuring a generational reduction in shader counts but relying on architecture and clock speeds for performance. It will have 10,752 CUDA cores spread across 84 SM (42 TPCs) and will retain the 256-bit bus width of the AD103, resulting in a bandwidth of 896 GB/s.

The GB205 will cater to the lower end of the performance segment with a surprising 6,400 CUDA core count spread across 50 SM in 5 GPCs, each with 5 TPCs. With a 192-bit memory bus width and a speed of 28 Gbps, this chip will offer 672 GB/s of memory bandwidth.

The GB206, aimed at the mid-range segment, closely matches the GB205's CUDA core count with 6,144 cores spread across 36 SM (18 TPCs) in 3 GPCs. The key differentiator is the memory bus width, narrowed to 128-bit for the GB206, resulting in 448 GB/s of memory bandwidth at the same 28 Gbps speed.

At the entry level, the GB207 features a significantly smaller chip with just 2,560 CUDA cores across 10 SM in 2 GPCs, each with 5 TPCs. The memory bus width remains at 128-bit, but the memory type reverts to the older GDDR6, with an expected bandwidth of 288 GB/s at 18 Gbps.

NVIDIA is anticipated to enhance all chips with larger on-die caches to bolster the memory subsystems. Innovations in ray tracing performance, AI acceleration, and other architecture-specific features are also expected. The "Blackwell" series is slated for a Q4-2024 launch, promising a significant leap in performance for gaming and professional markets.