Flagship GeForce RTX 5090 GPU 'GB202' to feature 192 SMs and 512-bit memory bus

Today, we've got some new info on NVIDIA's new 'Blackwell' GB202 GPU that will power the upcoming GeForce RTX 50 Series - and serve as the basis for the GeForce RTX 5090 much like the AD102 GPU served as the basis of the GeForce RTX 4090.

The specs come via reliable leaker and insider Kopite7kimi, who confirms that GB202 will sport a make-up of 12 GPCs and 8 TPCs that results in 192 SMs or Streaming Multiprocessors. For most PC gamers, this is technical jargon that won't make a lot of sense. Still, if we assume that the SM structure is similar to what we've seen with the Ada Lovelace generation, then GB202 - the fully unlocked GeForce RTX 50 Series GPU - will feature a whopping 24,576 CUDA Cores.

This represents a 33% increase over the AD102 Ada Lovelace GPU, and switching to TSMC's 3nm process should also lead to a more efficient GPU.

It's also worth mentioning that the GeForce RTX 4090 didn't use the full AD102 GPU, and currently, it's unlikely that we'll see a GeForce RTX 4090 Ti that does, so the eventual GeForce RTX 5090 could also end up using a cutdown version of the GB202 GPU with fewer SMs.

Regarding memory, Kopite7kimi also confirms that GB202 will feature a 512-bit bus - which adds weight to the rumors that NVIDIA will upgrade its GeForce lineup to GDDR7 for the next generation. Plus, with word that the RTX 5090 will boost up to 2.9 GHz out of the box and deliver 1.7X more performance than the RTX 4090, it will be a beast no matter how you slice it.

Of course, NVIDIA could make changes and still be in the process of locking down the hardware for the Blackwell range - which will see different GPUs being deployed for the workstation and AI market. These leaks also tell us very little about what sort of Tensor Cores, RT Cores, and other specialized hardware will be included - stuff that's proving to be a major part of NVIDIA's GeForce architecture outside of raw numbers.