The full-spec AD104 has 7680 CUDA cores: Probably called RTX 4070 Ti

There are rumors that in the first Ada Lovelace architecture GeForce RTX 40 series released by NVIDIA, the GeForce RTX 4070 is equipped with an AD104 GPU. It has 56 SMs, that is, 7168 CUDA cores, equipped with 10GB of GDDR6 memory, a bit width of 160 bits, and a rate of 18Gbps. While the specs don’t look high, the 3DMark Time Spy Extreme scores seem to be on par with the GeForce RTX 3090. However, in some cases, the bandwidth of the video memory is likely to become a performance bottleneck.
GeForce RTX 40 TSMC N5

Recently, Twitter user @kopite7kimi said that NVIDIA has also prepared a full-scale AD104 GPU with 60 SMs, that is, 7680 CUDA cores, equipped with 12GB of GDDR6 memory, a bit width of 192 bits, a speed of 21Gbps, and the corresponding PCB design is PG141-SKU331. Compared with the GeForce RTX 4070, its memory bandwidth is increased by 40%, and the overall performance can match the current GeForce RTX 3090 Ti.

Higher performance will come at the expense of higher power consumption, and the power consumption of a graphics card with a full-spec AD104 GPU reaches 400W, which is 100W higher than the GeForce RTX 4070. To help the name of the graphics card, according to Nvidia’s naming convention, it should be called GeForce RTX 4070 Ti.