A patent illustrating AMD’s plans to move to a chiplet or MCM design with its Radeon GPUs has been shared by Twitter user, Underfox. Although this is still just the proof of concept, it does mean that both AMD and NVIDIA will make the jump in the next 4-5 years. Let’s have a look:
This diagram immediately reminds me of Intel’s Ponte Vecchio block diagram. You’ve got both single-precision and double-precision compute units across multiple dies connected using Intel’s EMIB interconnect.
AMD will of course use its highly scalable Infinity Fabric, but the structure seems very similar. Team Red appears to be using four dies with two HBM2 memory stacks on each one. This will likely form an Exascale GPU powered by the CNDA architecture with double precision capabilities, and will likely complement the Zen 4 based Genoa server CPU. Like Intel, AMD likely won’t use stacking for its compute dies, but HBM stacking will most likely be leveraged as the block diagram indicates (3-Hi stacks).
Regardless, this means that the MCM design will first come to the Data Center and server space (unsurprisingly) as the bandwidth and compute requirement in this segment is vastly higher than that required by gamers.