New rumored specifications for three of NVIDIA’s upcoming GeForce RTX 40 Series graphics cards have surfaced today courtesy of kopite7kimi, providing potential insight at what early adopters might find when they unpack their Ada-based GPUs later this year. According to the popular leaker, the flagship GeForce RTX 4090 will feature 16,384 CUDA cores, an increased amount over earlier claims that suggested the premium graphics card would ship with 16,128 CUDA cores (the GeForce RTX 3090 Ti, for reference, features 10,752 CUDA cores). Kopite7kimi also shared the first alleged GPU specifications for the GeForce RTX 4080 and GeForce RTX 4070, which will purportedly feature 10,240 CUDA cores and 7,168 CUDA cores, respectively. These models may leverage GDDR6 memory instead of GDDR6X like their flagship GeForce RTX 4090 counterpart.
And TBP, 450/420?/300W.
— kopite7kimi (@kopite7kimi) June 23, 2022
This part of the [GeForce RTX 4080’s] specs is most likely to change though, because 18 Gbps GDDR6 memory would actually be slower than RTX 30 predecessors (RTX 3080 has 19 Gbps memory). However, one should underline the fact that AD102 GPU has larger L2 cache, which may work just as efficiently as Radeon RX 6000 series. In other words, the memory speed does not need to be that much faster.
The [GeForce RTX 4070] is no longer expected with 12GB memory but 10 GB instead. This would also mean a narrower memory bus from 192-bit to 160-bit. It is said that RTX 4070 would feature 7168 CUDA cores, which would mean 512 disabled CUDA cores on AD104 GPU.
Source: kopite7kimi (via VideoCardz)