AMD submitted a patent application to the US Patent and Trademark Office on December 31, 2020, showing a new modular GPU design method.
According to this patent, the new GPU will adopt MCM multi-chip modular design. Each GPU chip will have a dedicated area responsible for passive connection with adjacent cores. At the same time, the communication with the CPU will be handed over to the first GPU chip separately.
The four GPU chips will be packaged on the middle layer of the same PCB substrate. This design is similar to SoC chips and integrates many cores with different functions.
AMD said that traditional GPU design has inherent shortcomings, parallel load tasks are difficult to distribute among different GPU chips, memory access efficiency is also very low, it is difficult to maintain synchronization. In addition, most programs are only designed for a single GPU when they are written, and multi-graphics crossfire can be improved relatively small.
In addition, the current GPU chip area is getting larger and larger, leading to high manufacturing costs. Therefore, this patent can enable multiple GPU core chips to communicate better while reducing manufacturing costs.
Regarding the issue of memory synchronization, AMD stated that although each GPU chip has an independent last-level cache, the coupling method in the patent enables all small chips to run simultaneously. Since only the first GPU chip is needed to communicate with the CPU, from the perspective of the operating system and the CPU, it still works as a GPU.
This roadmap shows AMD’s plans. It can be seen that the MCM solution will integrate four GPU chips, and the small chip will have a dedicated channel to communicate with other chips. AMD’s next goal is to use modular technology to manufacture CPUs, that is, to manufacture I/O chips and CPU computing units separately, and to package chips of different processes together.
Although both Nvidia and Intel will launch their own MCM multi-chip solutions in the future, AMD has obviously made a lot of trials and accumulated experience in the previous Ryzen CPUs. The video cards said that AMD is expected to use this technology in RDNA3+ architecture graphics cards.
(Via)
