AMD has released a new patent that shows an active chiplet bridging multiple GPU cores, possibly based on an RDNA 3 architecture.
The chiplet design is something that has worked very well with AMD's Ryzen processors, and that seems to be the path the red company is going to choose for its next-gen GPUs.
AMD wants its next GPUs to have an MCM (Multi-Chip-Module) design and it is not the only one who is investing money and effort in it. Nvidia is also investing heavily to be able to adapt the MCM design to its upcoming GPUs.
The MCM design has proven to be very beneficial for AMD and will be crucial in the future to be able to design chips with a greater number of transistors without increasing the size of the chips too much. Increasing the density of transistors only by reducing the process nodes does not seem to be viable in the medium and long term, so large manufacturers such as AMD, Nvidia and even Intel have plans to use the MCM design in their next graphics cards. processors and APUs.
AMD claims that it faces a challenge to make multiple GPUs work in parallel. To address this issue, AMD has proposed an active bridge chiplet that would link multiple GPU chiplets together.
The main block diagram of the conceptual design shows a chip with multiple chiplets. The CPU part is connected to the first GPU chiplet (Next Generation Infinity Fabric), while the GPU chiplets are interconnected via the active bridge chiplet.
One of the most interesting revelations is the implementation of an L3 LLC (Last Level Cache) cache that is unified across the multiple chiplets. This helps reduce latency when reading the cache by being unified for all chiplets.
The diagrams indicate that this could be an SoC design for future APUs and RDNA 3 graphics, but implementation in future discrete GPUs and CDNA-based HPC products should also be expected. We will keep you informed.
0 Comments