NVIDIA shares full specs for AD102, AD103, and AD104 GPUs Ada Lovelace, which serve as the main graphics card GeForce RTX 4090, RTX 4080 16 GB and RTX 4080 12 GB, respectively. The company also named the range of GPUs and the number of transistors they have.
The AD102 GPU die area is 608 mm2 and it contains 76.3 billion transistors. The AD103, on the other hand, has a die area of 378.6 mm2, and it contains 45.9 billion transistors. The die area of the younger AD104 is 294.5 mm2and the number of transistors is 35.8 billion. For comparison, the flagship GA102 graphics chip of the Ampere generation has an area of 628 mm2 and contains only 28.3 billion transistors.
Among other things, the increased transistor budget served not only to increase the number of function blocks. They also allowed them to improve these blocks – NVIDIA redesigned CUDA cores, Tensor cores, and RT cores. In addition, NVIDIA has significantly increased the number of ROPs in the new GPUs compared to their predecessors. The AD102 chip has 190 ROPs (GA102 has only 112 ROPs), AD103 has 112 ROPs, AD104 has 80 ROPs.
Another part of the transistors was used to increase the amount of L2 cache in new GPUs. The flagship chip AD102 received 96 MB of L2 cache (GA102 had only 6 MB of L2 cache), AD103 – 64 MB and AD104 – 48 MB.
NVIDIA also corrected an inaccuracy in the description of the TSMC N4 production process based on the new Ada Lovelace GPUs. The company previously stated that this is a 4nm process, but in fact N4 is an advanced 5nm process.