New diagrams of the AMD GPU with chiplet design that emerged in the past months have surfaced. AMD seems to have made the idea of multiple chips even more efficient.
The red team is preparing to carry the multi-dice-based chip strategy it follows in Ryzen processors to the graphics cards.
Although the idea of a multi-membrane chip has significant performance and cost advantages, it also brings with it significant disadvantages. We saw this especially in the first 2 generations of Ryzen processors. However, AMD managed to address these negativities over time and offer more competitive solutions.
It seems that the company is in the preparation of applying the same design on graphics cards. After the multi-membrane design GPU diagrams shared in the past months, new drawings have emerged.
Unlike the first graphics, the new visuals contain an important update. In particular, the problems in communication between chips encountered in multi-membrane design and the delays experienced here worried after the first diagrams.
AMD aims to overcome this problem by locating a high-bandwidth active bridge with integrated memory instead of a passive bridge. Thus, while communication between chips will be provided, synchronization will also be strengthened.
Multiple dice will be able to work as a monolithic GPU
However, positioning the cache in the active bridge instead of a separate cache on each chip will allow different membranes to work practically like existing monolithic GPUs. Game developers will not need to make special arrangements for the chipset GPU.
According to the red team’s statements, this bridge will also be linked to Infinity Cache, which is currently available on RDNA2-architecture GPUs as its ultimate cache.
Finally, it is not yet known whether the diagrams have been prepared for CDNA or RDNA architecture chips. Adapting the graphic rendered lines of work to the chiplet design is not yet an achievable task. Time will show how the design without problems on the GPGPU wing will produce results.