by Antonio Delgado 2 hours ago …

AMD has published recently, just the 31 December, a new patent showing a chiplet-based GPU architecture.

The chiplet-based design divides a processor, or a GPU in this case, into different components manufactured separately and unified under a single interconnect layer. This system, already used in third generation AMD Ryzen processors for example, allows the design to be modularized and the same processor to use chiplets manufactured in different processes.

AMD will leverage its experience with the use of chiplets in its processors to bring them to GPUs

In the case of Zen 2 processors, for example, the CCDs where the cores are located are manufactured by TSMC at 7 nanometers, while the cIOD chiplet, commissioned of input / output and communication with the RAM memory through its own memory controller, it is manufactured to 12 nanometers in GlobalFoundries.

With this patent, AMD seeks to transfer this chiplet-based design to graphics cards, in such a way that modularized GPUs can be created with several smaller chiplets interconnected in one, being able to save costs when manufacturing (it is more It’s cheaper to create smaller chips and join them together than to create a bigger one).

The idea is that all the chiplets work as a single GPU, since, unlike with processors and the use of several cores in a conventional, widely supported operating system, multi-GPU systems depend a lot specific support for each game or application. For this, the patent shows an interconnection system that joins the different chiplets to a main chiplet, and this will be the one that communicates with the CPU. Each chiplet will have its own LLC cache.

Rumors seem to point to the first chiplet-based AMD graphics coming after the RDNA 3 architecture, an architecture that is scheduled for the year 2022, so it could be a new architecture or a evolution of RNDA 3 with a chiplet design.

End of Article. Tell us something in the Comments or come to our Forum!