Published in Graphics

AMD patents a new graphics-process construction method

by on20 June 2024

14 chiplets which function separately or together

AMD has secured a patent for an innovative GPU architecture that introduces a novel graphics-process construction method. This design features a total of 14 chiplets, organised in a way that allows them to function as a unified GPU or even as separate GPUs.

According to the Patent  the GPU is divided into three segments, each housing a front-end chiplet and three shader engine dies, which house the stream processors essential for rendering graphics. These clusters are interconnected and also linked to two additional dies—a mode selector and a multimedia die.

This concept bears a striking resemblance to the rumoured AMD Radeon Navi 4c chip, which was speculated to be part of AMD's upcoming RDNA 4 gaming GPUs like the Radeon RX 8900 XTX.

However, reports suggest that this particular design has been shelved, so this patent suggests that AMD has been exploring modular design.

Such an approach could offer significant manufacturing advantages. Producing complex monolithic dies is costly and yields fewer units per wafer. By breaking the GPU into smaller, discrete components, AMD could not only enhance production yields (as defects in a single chiplet are less impactful than in a whole GPU) but also have the flexibility to utilise different manufacturing processes for various parts of the GPU.

Similar to how AMD's Ryzen CPUs are fabricated using distinct processes for the CPU cores and the I/O hub, AMD could employ cutting-edge TSMC processes for the shader engine dies, while opting for older, less expensive methods for the multimedia and mode selector dies.

This modular strategy could also simplify AMD's product scaling. Rather than producing several unique monolithic dies, AMD could create a range of GPUs—from budget to high-end—using the same silicon base, differentiated only by the number of chiplets.

The patent abstract indicates that a variable number of GPU chiplets can be combined to form single or multiple GPUs, allowing for a diverse range of products to be created with minimal tape-outs. This method also supports the integration of chiplets from different technology generations.

The patent was filed on December 8, 2022, and published on June 13, 2024, suggests long-term planning has been involved. It remains to be seen if AMD will produce a multi-chiplet gaming GPU.

AMD’s plans so far have been to leave the high-end to Nvidia with RDNA 4 targeting the mid-range market.

Last modified on 20 June 2024
Rate this item
(4 votes)