A recently filed patent indicates that AMD may be shifting towards a chiplet approach rather than using monolithic designs for their upcoming GPUs.
AMD may soon be altering their direction, as most consumer CPUs and GPUs currently available on the market still utilize traditional monolithic chip architecture.
AMD has previously utilized the chiplet architecture, also referred to as a multi-chiplet module or MCM. One notable example is their Instinct MI200 AI accelerators, which employ an MCM design with multiple chiplets integrated into one package.
Despite this, the new patent suggests that AMD is prepared to incorporate chiplets into its mainstream RDNA architectures.
The primary focus of this patent is the utilization of MCM architecture in GPUs, and it showcases three distinct modes.
The initial mode is known as ‘single GPU’ mode, which would function in a similar manner to current GPU chips where all chiplets on the board would act as a unified processing unit. This may potentially enable compatibility with previous software and hardware.
The ‘independency mode’ is the second mode in which the chiplets can operate. In this mode, the chiplets are able to function independently, while a dedicated front-end die is responsible for managing task scheduling.
The last mode, known as ‘hybrid mode’, is potentially the most ambitious as it permits chiplets to function autonomously or collaboratively as needed.
Using this type of customizable MCM technology could provide advantages in scalability and resource utilization.
Despite current developments, AMD’s consumer-level graphics cards do not currently feature a chiplet-style architecture. While a chiplet-style GPU was initially planned for the RNDA 4 line-up, it has since been cancelled in favor of more traditional monolithic designs.
Perhaps AMD will adopt chiplet architecture for its upcoming RDNS 5 chips in order to continue improving performance, as the monolithic architecture is reaching a plateau in terms of potential performance gains.
Leave a Reply