NVIDIA 4nm AD104 “Ada” Silicon Pictured, Half the Die-area of AD102

Here’s the first picture of the 4 nm “AD104” silicon powering what would have been the $900 GeForce RTX 4080 12 GB, and upcoming RTX 4070-series graphics cards. The third largest GPU based on the “Ada Lovelace” graphics architecture, the AD104 looks tiny. This is because it has roughly half the die-area of the AD102, estimated to be around 295 mm² (compared to 608 mm² of the AD102), which means its transistor count should be less than half, with older reports pinpointing it to 35.8 billion. The RTX 4080 12 GB was supposed to max out the AD104 silicon, enabling all 60 streaming multiprocessors (SM) physically present.

The AD104 with 60 SM hence has 7,680 CUDA cores, 60 RT cores, 240 Tensor cores, 240 TMUs, and 80 ROPs. NVIDIA has generationally narrowed the memory interface (compared to the GA104 and TU104), down to 192-bit GDDR6X. Its predecessors such as the GA104 feature 256-bit wide memory interfaces. NVIDIA is overcoming the memory bus width “deficit” by giving SKUs based on the silicon higher memory speeds (21 Gbps or more); and architecture-level improvements such as larger on-die caches. NVIDIA is reportedly planning to launch an AD104-based SKU early January 2023. VideoCardz reports that could be the RTX 4070 Ti, a re-branding of the RTX 4080 12 GB.