Here is the first image of the 4 nm “AD104” silicon that will power the next RTX 4070-series graphics cards and the $900 GeForce RTX 4080 12 GB. The AD104, which is the third-largest GPU based on the “Ada Lovelace” graphics architecture, appears small. This is due to the fact that it has a die-area that is around half that of the AD102, estimated to be around 295 mm2 (as opposed to the AD102’s 608 mm2), meaning that its transistor count should be less than half that of the AD102, with earlier reports placing it at 35.8 billion. The AD104 silicon was intended to be fully utilized by the RTX 4080 12 GB, enabling all 60 streaming multiprocessors (SM) that were physically present.

Thus, the AD104 with 60 SM contains 80 ROPs, 240 Tensor cores, 60 RT cores, 7,680 CUDA cores, and 60 TMUs. In comparison to the GA104 and TU104, NVIDIA has reduced the memory interface generationally, reducing it to 192-bit GDDR6X. The GA104, one of its predecessors, has 256-bit wide memory interfaces. By offering SKUs based on the silicon faster memory speeds (21 Gbps or more), as well as architecture-level upgrades like bigger on-die caches, NVIDIA is addressing the “gap” in memory bus width. An AD104-based SKU from NVIDIA is supposedly set to debut in early January 2023. According to VideoCardz, that might be the RTX 4070 Ti, a rebranded version of the RTX 4080 12 GB.