AMD patents GPU chiplet designs, a future of RDNA architecture?

AMD seeks to say goodbye to monolithic chips, and while it has already done so with great success with its AMD Ryzen processors, now the next step is in its graphics cards, which will say goodbye to this type of design to advance to the chiplet, is that is, the incorporation of several graphics chips to offer great performance scaling. This is not a surprise considering that we have been many months since Nvidia and Intel already have graphics cards in development employing this design, so AMD is the last to join the party, although in its favor it has the experience accumulated during years with their desktop processors.

Diagram of a chiplet design on an AMD Radeon GPU

This information comes directly from AMD, who filed a patent with the United States Patent Office on December 31 for the potential design of a GPU with several graphics silicons under the same package. According to AMD, the monolithic designs of its GPUs have been maintained due to various implementation issues, including difficult workload distribution across multiple arrays (same issue already seen in Crossfire or SLI configurations). It is also complex and costly from a design point of view to synchronize the contents of memory across multiple graphics silicons.

AMD GPU diagram with chiplet design and high-bandwidth crossovers

To solve this problem, AMD has implemented the ” passive cross – linked high bandwidth ” (high bandwidth passive crosslinks). According to AMD, the first GPU chipset would be directly ” communicatively coupled ” to the CPU, while each of the GPU chips in the set would be coupled to the first GPU via a passive cross link. In this sense, AMD considered passive crossover as communication cables between chips that are placed in a single interposer (in several layers if necessary). Such a group of GPUs would function as a system on a chip, which is divided into different functional chips.

Diagram of an AMD GPU with chiplet design with illustration of the cache hierarchy

In conventional GPU designs, each of the graphics chips has its own top-level cache (LLC), but to avoid troublesome timing, AMD believes that each of the GPU chiplets should have its own LLC. , but in such a way that each of these caches is ” communicatively coupled ” to physical resources so that the cache remains ” unified and remains consistent across all GPU chiplets .”

Diagram revealing a GPU with a four-chiplet design

Although AMD has not made any official announcement that it is working on such a GPU, rumors indicated that the RDNA3 + architecture could be based on the chiplet design, while now this patent is only beginning to give credibility to the information. Also as we mentioned, Intel and Nvidia already have GPUs with multi-chip designs in development, so it is the next step for the industry to follow, and the first to do so will be Intel this year.


Back to top