The tech industry is constantly pushing boundaries to improve performance and efficiency, and AMD has been at the forefront with its chiplet-based designs. While their previous attempts with chiplet-based GPUs did not dethrone Nvidia’s dominance, a new patent suggests that AMD is not giving up on this innovative approach just yet. Dated November 23, 2023, the patent describes a radical departure from AMD’s existing chiplet layout. Instead of using one large main GPU die, the patent proposes a design that leverages multiple chiplets, including host memory cache dies (MCD) and a Graphics Compute Die (GCD).
By distributing the workload across several chiplets that work in parallel, AMD aims to achieve better performance and scalability. Interestingly, the patent indicates that no central chiplet will be responsible for distributing work to subordinates. Instead, each chiplet will function independently. While it remains unclear how exactly the chiplets will work autonomously without a central controller chip, this development reflects a potential shift towards chiplet-based designs in the industry. Monolithic dies have started reaching the limitations of lithography technologies, making chiplets a more viable option due to improved yields and reduced costs.
