AMD’s Patent Reveals a Unique “Multi-Chiplet” Approach for Future RDNA Architectures

AMD, a leading semiconductor company, has recently filed a patent that sheds light on its exploration of “multi-chiplet” GPU design options. This development suggests that next-generation RDNA architectures could feature significant changes, moving away from traditional monolithic designs.

The Rise of Multi-Chiplet Modules (MCMs)

While the concept of Multi-Chiplet Modules (MCMs) isn’t entirely new in the graphics industry, recent limitations with monolithic designs have led to a growing interest in MCMs. AMD, known for its expertise in multi-chiplet designs, has already implemented MCM solutions in its Instinct MI200 AI accelerators, stacking multiple chiplets on a single package. Additionally, the company introduced an MCM design in its recent RDNA 3 architecture with the Navi 31 GPUs.

Understanding AMD’s Patent

The patent outlines three distinct “modes” of chiplet utilization, each affecting how resources are allocated and managed:

Single GPU Mode: In this mode, all onboard chiplets function as a unified processing unit, sharing resources collaboratively. This mirrors the operation of modern-day GPUs.
Independency Mode: Individual chiplets act independently, with a dedicated front-end die responsible for scheduling tasks for associated shader engine dies.
Hybrid Mode: The most optimistic approach, where chiplets can act both independently and coexist. This mode combines the benefits of unified and independent processing, offering scalability and efficient resource utilization.

Challenges and Potential Benefits

While multi-chiplet configurations offer performance advantages and scalability, their production is more complex and requires high-end equipment and processes. The patent highlights the following benefits of the multi-chiplet approach:

Flexible Resource Configuration: By dividing the GPU into multiple chiplets, the system can flexibly allocate active GPU physical resources based on the operating mode.
Assembling Configurable GPUs: A configurable number of GPU chiplets can be assembled into a single GPU. This allows for constructing multiple GPUs with varying numbers of chiplets using a small number of tape-outs.

Consumer Segment and Future RDNA Architectures

Currently, AMD lacks a proper multi-GPU die solution for the consumer segment. While Navi 31 GPUs remain monolithic in design, the move to chiplet packaging for memory controllers and infinity cache hints at AMD’s future direction. We can expect next-gen RDNA architectures to embrace multi-chiplet packaging, potentially featuring multiple Graphics Compute Dies (GCDs) with dedicated Shader Engine blocks. Although a canceled RDNA 4 GPU under the Navi 4X/Navi 4C codename may not see the light of day, future RDNA 5 chips could bring further innovations.

Conclusion

With High-NA equipment and evolving technologies, the adoption of MCM designs may increase. AMD’s experimentation with multi-chiplets positions them well for a potential shift away from monolithic designs in future RDNA architectures.