According to the news of the core research institute on December 27, according to the news, the second half of next year will launch RTX 40 series graphics cards, using TSMC 5nm process. NVIDIA's next-generation GPUs include Two architectures, Hopper and Ada Lovelace, both based on TSMC's 5nm process, of which Hopper is mainly oriented to data centers, serving AI, cloud computing, metaversities and other scenarios, multi-new MCM design, and will use TSMC's CoWos fan-out package, which is expected to tape out soon.
Edited by the Institute of Core Research
Ada Lovelace is a game graphics card GPU, corresponding to the RTX 40 series, of which the AD102 core is expected to integrate as many as 18432 CUDA cores, an increase of more than 71% over the ampere architecture GA102, plus the frequency will reach 2.2GHz, theoretically FP32 floating-point computing performance can reach 81TFlops, compared to RTX 3090 32TFlops increased by 1.25 times. The RTX 40 series has been greatly upgraded, and NVIDIA's manufacturing process will also shift to TSMC 5nm, but they are no longer TSMC's priority customers, and the priority lags behind Apple, MediaTek, AMD and other manufacturers, and it is not simple to obtain stable 5nm production capacity.
Industry sources said that TSMC has low requirements for apple, MediaTek, AMD and other three major customers, and does not need to prepay too much margin to stabilize production capacity, and customers such as NVIDIA need to pay a huge advance payment in advance if they want to obtain 5nm production capacity orders.
According to NVIDIA, they advanced $1.64 billion in Q3 quarter and will pay $1.79 billion in the future, while the entire long-term order advance was as high as $6.9 billion, well above last year's level. The $6.9 billion order is uncertain about the ratio of Samsung to TSMC, but from the perspective of process upgrade, finding TSMC to buy 5nm production capacity accounts for the majority.
(7839069)
![The RTX40 graphics card uses TSMC's 5nm process and Nvidia spent $6.9 billion on it[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)