AMD's patent filing reveals that the company is exploring "multi-chiplet" GPU design options, suggesting that next-gen RDNA architectures might feature massive changes.
AMD Looks Towards Shifting Away From Monolithic Designs Through The Adoption of Multi-Chiplet Approach In Next-Gen RDNA GPUs
The concept of MCM (Multi-Chiplet Module) isn't entirely new for the graphics segment, but with the limitations of monolithic designs, the inclination towards MCMs in the industry is surely growing.
AMD looks to be one of the firms which are well-experienced with multi-chiplet designs, since their lineup of Instinct MI200 AI accelerators was the first to feature an MCM design with multiple chiplets stacked on a single package, such as the GPCs (Graphics Processing Cores), HBM stack and the I/O die. The firm was also the first to feature an MCM solution on its most recent RDNA 3 architecture with the likes of Navi 31. However, with the new patent, Team Red looks to translate this ideology into mainstream "RDNA" architectures, and here's how.
The patent describes the use of three different "modes" of chiplet utilization, where the difference lies in how resources are allocated and then managed ahead. The patent reveals three different models, with the first one being the "single GPU" mode, which is quite similar to how modern-day GPUs function. All the onboard chiplets will act as a single, unified processing unit, with resources shared in a collaborative environment.
The second mode is called the "independency mode", where individual chiplets would act independently, responsible for their functions through the dedicated front-end die responsible for scheduling tasks for its associated shader engine dies. The third mode is the most optimistic of them all and is called the "hybrid mode" where chiplets can act independently and co-exist as well. It leverages the benefits of both unified and independent processing, offering scalability and efficient resource utilization.
The patent doesn't reveal details regarding's AMD approach with their MCM design so we can't comment on whether Team Red would decide to adopt the ideas mentioned in the patent. However, talking about multi-chiplet configurations in general, while they do offer performance leverages and scalability, producing them is a much more complex task and it requires high-end equipment and processes, ultimately increasing costs as well. Here's how the patent describes the multi-chiplet approach:
By dividing the GPU into multiple GPU chiplets, the processing system flexibly and cost-effectively configures an amount of active GPU physical resources based on an operating mode.
In addition, a configurable number of GPU chiplets are assembled into a single GPU, such that multiple different GPUs having different numbers of GPU chiplets can be assembled using a small number of tape-outs and a multiple-die GPU can be constructed out of GPU chiplets that implement varying generations of technology.
Currently, AMD doesn't have a proper multi-GPU die solution for the consumer segment. The Navi 31 GPUs are still very much a monolithic design with a single GCD but the MCDs which carry the infinity cache and memory controllers have been moved to a chiplet package. With the next-gen RDNA architectures, we can expect AMD to go even heavier on the multi-chiplet packaging with multiple GCDs having their own dedicated Shader Engine blocks. AMD had one such GPU planned for the RDNA 4 lineup under the Navi 4X/Navi 4C codename but that has reportedly been canceled in favor of more mainstream monolithic packages so maybe we can see it return with the future RDNA 5 chips.
With High-NA equipment and rapidly evolving technologies in play, the adoption of MCM designs could increase as well, and given that Team Red has already experimented multi-chiplets, it's certainly not of the equation if we talk about future RDNA architectures shifting away from monolithic designs.
News Source: Tom's Hardware
