Nvidia’s heavy cost to supply TSMC 5-nanometer wafers in the new generation


Apparently Nvidia to pre-purchase the 5-nanometer TSMC wafers Used in the new generation of GPUs GeForce RTX 40 It has cost a lot.

As you probably know, the new generation of Nvidia graphics chips called Ada Lovelace, the graphics processing power of GeForce RTX 40 series products And based on the information we have received so far, it is supposed to be in the architecture of this GPUs from TSMC 5nm lithography to be used. Both companies AMD and Nvidia are using this production process in their new generation of graphics products, but it seems that Nvidia is determined to make sure that the wafers it uses are sufficient. To that end, they paid several billion dollars in advance to buy the 5-nanometer wafer to the Taiwanese manufacturer of semiconductor chips.

According to the MyDrivers website, Nvidia has already paid $ 1.64 billion to TSMC in the third quarter of 2021 and is expected to transfer another $ 1.79 billion to the company’s account in the first quarter of 2022. The multi-billion dollar deal between the two companies will eventually cost the Green Team $ 6.9 billion, far more than they spent last year. In fact, in addition to TSMC, Nvidia has taken the same approach to Samsung, but it seems that a large part of these costs are due to TSMC 5-nanometer technology. Will be allocated.

NVIDIA GeForce RTX 4090 graphics card, Ada Lovelace flagship generation based on AD102 graphics processor

RTX 30 graphics card

According to previous rumors, the new generation of Nvidia GPUs based on the Node N5 process Company TSMC Includes flagship model The AD102 is based on the high-end GeForce RTX 4090 graphics card It is used and has a completely integrated design. It is said to be the maximum operating frequency of the graphics chip AD102 It reaches 2.5 GHz, the average value of which will be 2.3 GHz in boost mode. However, in the exact text of this tweet, which was published by an unofficial source, a figure of 2.3 GHz or higher is mentioned, which is currently considered as a forecast criterion.

It has been said so far that the AD102 chip From 18432 cores CUDA Of course, this figure can still be changed, but the same number of computing cores in the form of 144 units SM Located and approximately twice the number of cores used in the flagship series Ampere is still a big leap forward for the Turing generation It was considered. Speed ​​2.3 to 2.5 GHz This processor theoretically leads to 85 to 92 teraflops of graphics processing power, which is more than twice the performance of 36 teraflops of powerful RTX 3090 graphics card in the current generation.

Nvidia RTX graphics card

The 150 percent improvement in Nvidia GPU performance is a big leap forward, but keep in mind that Nvidia also experienced a significant improvement over the previous generation RTX graphics card. The flagship GA102 chip used in the RTX 3090 graphics card has 36 teraflops of graphics processing power. While the TU102 chip in the flagship RTX 2080 Ti graphics card provided only 13 teraflops of graphics processing power.

This indicates that the Ampere family was theoretically more than 150 percent stronger than the Turing series, when in fact the difference was as high as 50 to 60 percent. So we have to keep in mind that the number of teraflops these days does not directly indicate the performance of a GPU. In addition, we do not know if the 2.3 to 2.5 GHz value corresponds exactly to the card’s average clock speed or to its maximum boost speed, indicating that the AD102’s computing power potential could be higher.

In addition, the whistleblower claims that the NVIDIA GeForce RTX 40 series flagship, similar to the RTX 3090, has a 384-bit bus interface. Interestingly, the G6X is mentioned here, which means that Nvidia will not go for the new memory standard in the next generation, to use a newer technology (such as GDDR7) for this purpose after the Ada Lovelace series. This card uses 24 GB of memory, which is to be expected from a 16 GB one-way DRAM module or an 8 GB DRAM dual module.

GeForce RTX 40 graphics card

Comparison of Nvidia GPUs from generation to generation

GPU name

TU102 GA102 AD102
Architecture Turing Ampere Ada Lovelace
Manufacturing technology TSMC 12nm NFF Samsung 8nm 5nm
Graphic Processing Clusters (GPCs) 6 7 12
Texture Processing Clusters (TPCs) 36 42 72
Current multiprocessors 72 84 144
Number of CUDA cores 4608 10752 18432
Traflops Theory 16.1 37.6 90
Memory type GDDR6 GDDR6X GDDR6X
Memory bus 384bit 384bit 384bit
Memory capacity 11GB (2080 Ti) 24GB (3090) 24GB (4090?)
Flag model RTX 2080 Ti RTX 3090 RTX 4090
Power 250W 350W 450-650W
Release time Sep. 2018 Sept. 20 2022

Nvidia’s Ada Lovelace GPUs in the new generation of GeForce RTX 40 graphics cards will compete directly with AMD’s RDNA 3 architecture-based chips in the Radeon RX 7000 graphics card. There are still speculations that Nvidia may use the MCM architecture style. The MCM-structured Hopper GPU has been developed specifically for data centers and AI-based equipment, and their design process will be completed soon. Of course, it is said that Nvidia is currently limiting the MCM design to these products and will not use it in the Ada Lovelace series.

Finally, you can share your thoughts about this with Zomji and other users.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *