AMD Patents Chiplet Architecture for Radeon GPUs

On December 31st, AMD’s Radeon group has filed a patent for a chiplet structure of the GPU, displaying its imaginative and prescient about the way forward for Radeon GPUs. Currently, the entire GPUs out there in the marketplace make the most of the monolithic method, which means that the graphics processing models are situated on a single die. However, the present method has its limitations. As the dies get greater for high-performance GPU configurations, they’re costlier to fabricate and cannot scale that effectively. Especially with fashionable semiconductor nodes, the prices of dies are rising. For instance, it will be extra economically viable to have two dies which might be 100 mm² in dimension every than to have one at 200 mm². AMD realized that as effectively and has thus labored on a chiplet method to the design.

AMD studies that the usage of a number of GPU configuration is inefficient attributable to restricted software program help, so that’s the reason why GPUs have been saved monolithic for years. However, it looks like the corporate has discovered a technique to go previous the constraints and implement a adequate resolution. AMD believes that by utilizing its new excessive bandwidth passive crosslinks, it may possibly obtain splendid chiplet-to-chiplet communication, the place every GPU within the chiplet array can be coupled to the primary GPU within the array. All the communication would undergo an energetic interposer which might comprise many layers of wires which might be excessive bandwidth passive crosslinks. The firm envisions that the primary GPU within the array would communicably be coupled to the CPU, which means that it should use the CPU probably as a communication bridge for the GPU arrays. Such a factor would have massive latency hit so it’s questionable what it means actually.

The patent additionally suggests that every GPU chiplet makes use of its personal Last Level Cache (LLC), as a substitute of separate LLCs for every GPU, so every of the LLCs is communicably coupled and the cache stays coherent throughout all chiplets. Rumors recommend that we’re going to see the primary chiplet-based structure from AMD as successor to the RDNA3 era, so it would occur within the coming years. AMD already has expertise with chiplets from its processors, with Ryzen processors being the prime instance. We simply want to attend and see the way it will look as soon as it arrives for GPUs.