Nvidia Lovelace GPU architecture will use TSMC’s N5 process

In the past period of time, some people in the industry have further confirmed the information of Nvidia Ada Lovelace architecture. As Nvidia’s next-generation game graphics card for players, the news has always been confusing, and the information leaked from different channels is also different.

Recently, Twitter user @kopite7kimi confirmed again that NVIDIA Ada Lovelace GPU architecture will be manufactured using TSMC’s N5 process. Since this person predicted almost the entire Ampere architecture lineup of Nvidia a few months in advance, news about Nvidia should be believed. Previously, people from all parties have debated for a long time whether to choose TSMC or Samsung for the Ada Lovelace GPU architecture. In addition, @kopite7kimi also said that Nvidia will release the Ada Lovelace GPU architecture sooner than expected. In fact, someone said before that the Ada Lovelace GPU architecture will be launched before the end of 2022, probably between the third quarter and the fourth quarter of 2022.
Unlike AMD’s use of multi-chip module design (MCM) on the RDNA 3 architecture, NVIDIA does not plan to do so on the GeForce RTX 40 series. In the Ada Lovelace GPU architecture, the flagship product AD102 is a large 600mm² chip. It is rumored to be equipped with 18,432 CUDA cores. The performance is doubled on the basis of GA102. The power consumption is about 400W-450W. AD102 equips with GDDR6X video memory, the bit width is 384 bits. Except for AD102, there are not many news about AD104 or AD106.

However, Nvidia also has chips designed with MCM. It is rumored that it is a Hopper GPU architecture used in data centers. It will be Nvidia’s first product with this design. It will be manufactured based on TSMC’s 5nm process. Its GPU will have two chips, a total of 288 SMs, which can provide three times the performance of the current A100 computing card. The new computing card equipped with this GPU may become an H100.