NVIDIA RTX 50 “GB202” Gaming GPU To Feature The Same TSMC 4NP Process As B100

Yesterday, NVIDIA revealed its newest Blackwell GPU architecture tailored for data centers and AI acceleration, signaling the initial phase of their architectural advancement. According to reports, the company plans to introduce a new range of gaming GPUs later this year, all based on the Blackwell framework. Notably, the upcoming B100 GPU will maintain proximity to its predecessor in terms of process nodes, opting for 4NP, an improved version of 4N, a custom TSMC node initially crafted for the Hopper series.

Per Kopite7kimi, NVIDIA will maintain the status quo with their gaming GPU series, sticking to this node. As per the statement, this fresh node is anticipated to deliver a 30% boost in transistor density. Consequently, both Ada and Blackwell will adopt a variant of the 5nm node.

The leaker mentioned that the Blackwell GB202 gaming GPU will undergo a modification in its L1 cache (distinct from L2 cache), expected to bring notable enhancements compared to AD102 and GA102, both equipped with a 128 KB size. This alteration could potentially boost the throughput of a single Streaming Multiprocessor, according to Kopite7kimi.

images 1

The GB202 GPU is anticipated to serve as the flagship processor for the RTX 50 series. Speculation suggests it could feature 192 Streaming Multiprocessors (SMs) and a potentially robust 512-bit memory bus. Considering the dual-die design of the B100, a similar approach might be adopted for the smaller GB202 gaming processor. Additional rumors indicate that GB203 specifications could be halved compared to GB202, possibly encompassing around 96 SMs and half the memory bus width.

Details regarding the specs of the other Blackwell gaming GPUs remain unclear. However, it was mentioned that all RTX 50 series GPUs should support PCIe Gen5, DisplayPort 2.1, and incorporate onboard GDDR7 memory.