It is rumored that NVIDIA will release GPUs with Ada Lovelace architecture at the end of this year, and the flagship models will use AD102, including GeForce RTX 4080/4090 series graphics cards. It is said that the
GeForce RTX 4090 will have up to 144 groups of SMs, that is, 18432 CUDA cores, with 24GB of GDDR6X memory, a bit width of 384 bits, and a 12+4Pin new 12VHPWR power supply interface, and the power consumption will reach 600W or more. It is rumored to have a clock of 2.2 GHz, providing 81 TFLOPs of computing performance (FP32).
Recently, Twitter user
@kopite7kimi revealed that Nvidia has begun testing AD102, using a memory chip with a rate of 24 Gbps. While it’s impossible to verify that the content is accurate, based on the flow and timing of general GPU production testing, this information is logical if Nvidia really intends to release GeForce RTX 40-series graphics cards around this fall.
Unlike the current GeForce RTX 30 series GPUs based on the Ampere architecture, Nvidia chose TSMC’s 5nm process node on the Ada Lovelace architecture GPU, replacing Samsung as the foundry for the new generation of chips.
According to reports, Nvidia has prepaid TSMC $1.64 billion in the third quarter of 2021 and paid another $1.79 billion in the first quarter of 2022, advance payments for the entire long-term order will reach $6.9 billion to secure chip production. Previously, at GTC 2022, NVIDIA released a new generation of H100 based on the Hopper architecture, manufactured using TSMC’s 4nm process, and a version tailored for NVIDIA. It can be seen that Nvidia has also made great efforts in terms of technology and production capacity.