AMD Explores Multi-Chiplet GPU Design with Patent

TapTechNews June 15th news, AMD's newly approved technical patent shows that the company is exploring a multi-chiplet (multi-chiplet) GPU design scheme, This indicates that the next-generation RDNA architecture may undergo tremendous changes.

The multi-chiplet module (MCM) is not a new concept. However, due to the increasingly prominent limitations of single-chip designs, the industry is increasingly prone to MCM solutions.

AMD also has rich experience in MCM design. For example, the Instinct MI200 AI accelerator series took the lead in adopting MCM design, stacking multiple chiplets such as graphics processing cores (GPC), HBM stacks, and I/O chips on a single package. AMD also took the lead in adopting the MCM solution on its latest RDNA3 architecture such as Navi31.

AMD's patent describes three different modes, which differ in how resources are allocated and managed. TapTechNews attaches the relevant content as follows:

Uniform GPU

This is similar to the mainstream GPU function. All on-board chips will serve as a unified processing unit and share resources in a collaborative environment.

Independent mode

In this mode, each chiplet operates independently and is responsible for scheduling tasks for its associated shader engine chip through a dedicated front-end chip.

Mixed mode

In this mode, the chiplets can operate independently or coexist. It takes full advantage of the advantages of unified processing and independent processing and provides scalability and efficient resource utilization.