AMD has big plans for the future of consumer GPUs, and those plans are nothing short of revolutionary. The company, known for its bold innovations, is working towards adopting multi-chiplet GPUs, a move that could drastically change the GPU landscape. While the concept of Multi-Chiplet Modules (MCM) isn’t new to the tech world, its use in GPUs has been limited by the challenges of latency and data transfer. 
But AMD, with its extensive experience in chiplet designs, is looking to push the boundaries of what’s possible in gaming GPUs.
One of AMD’s biggest successes in this area has been its Instinct MI200 series of AI accelerators, which were the first to implement an MCM design with multiple chiplets stacked together in a single package. This design includes Graphics Processing Cores (GPCs), HBM stacks, and an I/O die. The upcoming Instinct MI350 lineup is taking this design further, offering a glimpse of what might be in store for consumer GPUs.
The main challenge with adopting chiplet-based GPUs for gaming lies in the latency. With frames needing to be rendered quickly, long-distance data hops can be detrimental. AMD’s solution seems to come from a new patent that reveals how the company plans to use a data-fabric circuit with a smart switch to bridge the communication gap between compute chiplets and memory controllers. This switch, essentially an optimized version of AMD’s Infinity Fabric, will allow for faster and more efficient data transfer between chiplets, addressing the latency issue that has plagued previous designs.
The new architecture will feature GCDs (Graphics Compute Dies) with both L1 and L2 caches, similar to what’s found in AI accelerators. Additionally, a shared L3 cache, or stacked SRAM, will connect all the GCDs, reducing the need to access global memory and allowing chiplets to communicate faster with each other. AMD’s approach is aimed at creating a system where the data exchange between chiplets is as fast as possible, ensuring smooth performance for gaming applications.
One of the most exciting aspects of this patent is how it reveals AMD’s readiness to implement these chiplet designs in consumer GPUs. The company has already worked on ecosystem-ready solutions, such as using TSMC’s InFO-RDL bridges and Infinity Fabric between dies for packaging. This level of readiness suggests that the transition to multi-chiplet consumer GPUs could happen sooner than expected. Moreover, with AMD’s UDNA architecture already set to merge gaming and AI, the potential for innovation in the GPU space is greater than ever.
However, the road to multi-chiplet gaming GPUs isn’t without its challenges. Latency is still a major concern, and even though AMD’s new switch mechanism promises to reduce it, there are still risks involved. For now, enthusiasts will have to wait and see whether AMD can overcome these hurdles and bring multi-chiplet GPUs to the gaming market.