This new CoWoS technology will bring the maximum value of NVIDIA graphics cards, among which a very small size of GPU and important reduce electricity consumption, which in turn produces less thermal births and, therefore, may reduce the need to integrate excessive heatsinks to cool.
NVIDIA and AMD are already using CoWoS
NVIDIA has already used this CoWoS design in the past on some of its cards graphics dedicated to the professional field, but only on high-end GPUs like Titan, Quadro and Tesla. This technology was, in fact, based on the Pascal era, and the GP100 (Pascal) and GV100 (Volta) GPUs were already developed using the TSMC CoWoS system. The first was based on the 16 nm FinFET process, while the second benefited from the 12 nm process.
AMD Vega 20 GPUs produced at 7 nm use TSMC CoWoS technology, and the source has not mentioned it in its report. This means that NVIDIA, Xilinx and HiSilicon will be the three largest TSMC customers in the area, and since the production capacity is between 6,000 and 8,000 per month, it makes sense that this is because It can only come to cover the need to build one customer
Currently only to receive technical drawings
Unfortunately, it seems that we will not see this NVIDIA graphics project being launched with Chiplets without a professional component, at least for now. This is due to the high production cost of these kits, and that NVIDIA will have an impact on the price that consumers will pay for graphics cards, which is much cheaper (remember that today the RTX 2080 Ti still costs more than 1,300 euros).
Therefore, we repeat that initially, the design of the Chiplets using the TSMC CoWoS process in NVIDIA drawings would be limited only to the professional sector, with Quadro and Tesla drawings similar to the previous one. There is no indication that NVIDIA intends to extend the project to Ampere, but are likely to start using MCM (Multi-Chip-Module) technology for the next generation of graphics cards, Hopper.
Recall that a few days ago rumors surfaced about three NVIDIA nameless graphics card models, but supposedly members of a professional range, which consisted of dozens of CUDA calls and VRAM memory. Of course, these are probably the first NVIDIA GPUs for the project we talked about today.