One of the most famous NVIDIA GPU leakers has just confirmed key details about the company's upcoming next-generation Ada Lovelace GPU. The new architecture will be built on TSMC's 5nm process. If it is true, this will be big news-let's discuss it first. Kopite leaked almost all the details of Turing and Ampere's architecture and was historically very accurate. In other words, as before the leak and multiple confirmations, a grain of salt will never hurt anyone.
Kopite has previously confirmed that the Ada Lovelace architecture will be built on a 5nm process, but it is assumed that this will be a Samsung node. Nvidia made the (correct) decision to transfer its foundry service to Samsung because it (accurately) predicted TSMC's demand. Having to share resources with Apple and AMD is a gamble, and NVIDIA's decision to cooperate with Samsung seems to have been mostly verified.
— Kopite7kimi (@kopite7kimi) August 26, 2021
However, if this leak is correct, then NVIDIA will return to TSMC's 5nm process-which may mean that TSMC will produce at least one flagship GPU series-which is a business that Samsung foundries have just lost. Even if they manufacture high-end chips at TSMC and keep low-end chips at Samsung (which is where sales are usually located), it is still interesting because it means that Jensen has lost confidence in Samsung's ability to keep up with the most cutting-edge nodes.
Ada Lovelace will come out earlier. keep patient.
— Kopite7kimi (@kopite7kimi) August 26, 2021
Kopite also stated that the Ada Lovelace GPU will arrive earlier than expected, although since they have not confirmed the schedule first, we still don't have a specific date. There are rumors that the Ada Lovelace GPU will arrive between the second quarter of 2022 and the fourth quarter of 2022, so if Kopite expects NVIDIA to postpone production and start mass production a little earlier, the timetable for the second quarter of 2022 may be the future Timetable to remember.
In many ways, Ada Lovelace can be considered the world's first computer enthusiast. She was the first person to realize that the analysis engine proposed by Charles Babbage has applications beyond pure computing, and she also published the first algorithm believed to be carried by such a machine (becoming the first computer programmer). It has been almost half a century since Alan Turing completed their work and invented the general-purpose computer during the World War.
As we all know, NVIDIA's architecture is based on famous physicists, mathematicians and scientists, and Ada Lovelace is no exception. Videocardz actually managed to find a major clue in NVIDIA's own merchandise store, which seems to confirm the rumor that the Lovelace architecture is the company's next-generation GPU. If you look at the heroes shown in the GTC 2018 keynote, you will not only find Ada Lovelace, but also all possible architectural code names for NVIDIA in the future. Jensen may have secretly left the entire future roadmap (in terms of code name) during the GTC'18 keynote.
According to the preliminary specifications provided by Kopite (which can be changed), the NVIDIA AD102 "ADA GPU" seems to have 18432 CUDA cores. This is almost twice the core of Ampere, and Ampere is already a big improvement for Turing. The only way this is even possible is because NVIDIA is clearly built on a 5nm process, which has a significant chip area and power consumption reduction. Interestingly, if you assume a clock speed of 1.75 GHz, you can also get the peak single-precision performance of the ADA 102 GPU: 64 TFLOP.
According to Kopite, the ADA architecture will have a larger L2 cache (Turing and Ampere both have a 6 MB cache), which means this may be a major architectural revision (because Turing is Pascal and Pascal is Fermi/ Kepler), not just the usual process shrinkage.
The latest technology news in your inbox, once a day!