Korean memory manufacturer SK Hynix and Nvidia are going to collaborate on an innovative gpu redesign, several South Korean publications, including The Joongang, write. They reportedly plan to stack High-Bandwith Memory (specifically HBM4) directly on top of processors.
The closer HBM can be placed to the gpu chips, the more efficiency can be achieved. Until now, HBM stacks (which consist of multiple memory devices, often between eight and sixteen, and a logic layer) were placed on an interposer next to the gpu’s or cpu’s and connected to the processor with a 1024bit interface. However, by integrating these stacks instead directly on to the processor versions, interposers are no longer needed and efficiency can be increased. This makes it very similar to AMD’s 3D-V-Cache technology, although HBM4 is expected to deliver higher capacities and be cheaper, but presumably slower.
It is not yet entirely clear exactly how the companies intend to pull this off. SK Hynix is reportedly still testing the exact method, including with manufacturers other than Nvidia. Nevertheless, it is expected that the US chip manufacturer will be the official partner, and that this new gpu architecture will be manufactured with TSMC’s wafer bonding technology. HBM4, by the way, is still under development, and instead of a 1024bit memory bus (as in previous HBM iterations), it is to include one of 2048 bits, which greatly increases its complexity.