AMD patents indicates company is embracing chiplet design
A new patent suggests AMD might be moving away from monolithic designs and towards a chiplet model for its future GPUs.
Though most consumer CPUs and GPUs currently on the market use traditional monolithic chip architecture, AMD might soon be changing direction.
Chiplet architecture, more properly known as multi-chiplet module or MCM, has been used by AMD before. Notably, their line-up of Instinct MI200 AI accelerators used an MCM design that features multiple chiplets in a single package.
The new patent, however, indicates AMD is ready to bring chiplets into its mainstream RDNA architectures.
The patent seems focused on using MCM architecture in GPUs and features three different modes.
The first is ‘single GPU’ mode, which would work similarly to existing GPU chips with all onboard chiplets acting as a single unified processing unit. This could allow for backward compatibility with older software and hardware.
AMDThe second mode is referred to as the ‘independency mode’. This would allow the chiplets to act independently, while a dedicated front-end die would be responsible for scheduling tasks.
The final mode is possibly the most ambitious and is labeled as ‘hybrid mode’, which allows chiplets to either act independently or cooperatively as demand requires.
The use of this style of configurable MCM technology could offer benefits in terms of scalability and utilization of resources.
However, AMD currently does not offer any chiplet-style architecture on its consumer-level graphics cards. A chiplet-style GPU was reportedly planned for the RNDA 4 line-up at one point, but this has supposedly been canceled in favor of more traditional monolithic designs.
Monolithic architecture is approaching a plateau where it will be difficult to squeeze further performance gains. Perhaps AMD will adopt chiplet architecture for its upcoming RDNS 5 chips to overcome this.