Please note that this post is tagged as a rumor.
NVIDIA AD102 GPU to feature more than 75B transistors
According to Kopite7kimi, NVIDIA new flagship GPU called AD102 is to feature 75 billion transistors, 2.65 times more than GA102.
With 75 billion transistors, AD102 GPU would only have 5B fewer than Hopper H100 GPU, NVIDIA’s flagship data-center processor based on the same TSMC 4N node. The size of the NVIDIA AD102 GPU is still unknown, a direct comparison between Hopper and Ada is not possible presently, but both processors are to feature a large L2 cache (50MB for Hopper and up to 96MB for Ada). Almost twice as large AD102 GPU may explain where those transistors were used.
Kopite has not confirmed the exact transistor count, but he did confirm it is more than 75B. The exact number may not leak until the official announcement exactly next week at GTC.
Remember this. 😁Most people made the right choice. More than 75B.
— kopite7kimi (@kopite7kimi) September 13, 2022
NVIDIA AD102 GPU is to feature up to 18432 CUDA cores, but probably not for the first wave of GeForce RTX GPUs. This processor is to be offered with GDDR6X memory and supports up to 384-bit bus width. Thus far only one model is known to feature this GPU, the RTX 4090 which is said to have 16384 cores enabled. Rumors suggest though that NVIDIA will be launching models with more cores, such as RTX 4090 Ti or possibly even a TITAN.
|NVIDIA Flagship Gaming GPUs|
|Node||TSMC 4N||SAMSUNG 8N||TSMC 12nm||TSMC 12nm||TSMC 16nm|
|Die Size||TBC||628 mm²||754 mm²||815 mm²||471 mm²|
|Trans. Density||TBC||45.1M TRAN/mm2||24.7M TRAN/mm2||25.9M TRAN/mm2||25.5M TRAN/mm2|
|Memory Bus||GDDR6X 384-bit||GDDR6X 384-bit||GDDR6 384-bit||HBM2 3072-bit||GDDR6X 384-bit|