SK Hynix is reportedly hiring design workers for logic units, comparable to CPUs and GPUs, to develop a stacking technique to place HBM4 immediately on prime of them. This would revolutionize the foundry sector by altering how logic and reminiscence elements are coupled and manufactured. As per the identical report, Nvidia appears on this method and is working with SK Hynix on its growth.
According to The JoongAng (through Tom’s Hardware), the corporate hopes to get rid of the necessity for interposers by putting HBM4 stacks immediately on CPUs, much like AMD’s direct-to-CPU 3D V-Cache implementation. While HBM can be costlier than cache, it would have a bigger capability, making it a less expensive answer.
Image credit score: Joongang.co.kr
HBM stacks mix a hub-like logic layer with a number of reminiscence units related to CPUs or GPUs through a 1024-bit interface. However, SK Hynix’s technique of mounting HBM4 stacks immediately on CPUs will cut back prices and simplify chip designs. The firm is claimed to be in talks with fabless companies, together with Nvidia, about their HBM4 built-in design method. The two corporations are anticipated to collaborate to construct the chip from the bottom up and have it produced by TSMC, which may even use wafer bonding to connect SK Hynix’s HBM4 machine to logic chips.
However, there may be one challenge that must be addressed: thermals. GPUs like Nvidia’s H100 have excessive energy consumption and warmth dissipation, and HBM reminiscence can also be identified for being extra power-hungry than different reminiscence options. More advanced cooling methods, comparable to liquid cooling or submersion, could also be wanted to chill down a bundle containing logic and reminiscence.
According to Professor Kim Jung-ho of KAIST’s Department of Electrical and Electronics, HBM and GPU can perform as one unit with out an interposer, however provided that the heating drawback is solved, which ought to occur in two to a few generations. The direct integration of reminiscence into CPUs may even alter the structure and manufacturing means of chips and will impression costs too.
KitGuru says: Processor architectures have at all times separated reminiscence and logic semiconductors. However, if SK Hynix can discover a viable technique to mass-produce chips with 3D-stack HBM, the foundations could also be about to vary comparatively quickly.
