SK Hynix and Nvidia are working on HBM4 modules with a revamped design, stacking 3D memory directly onto chip cores.
HBM4: SK Hynix and Nvidia Aim to Integrate this Memory in Future GPUs
SK Hynix could develop an HBM4 memory where modules are stacked directly onto the GPU. This collaboration with Nvidia aims to implement it in their upcoming GPUs in the future.
In the current scenario, HBM stacks integrate 8, 12, or 16 memory devices, along with a logical layer acting as a hub. These HBM stacks are placed on the interposer alongside CPUs or GPUs and connect to their processors through a 1024-bit interface. The idea with HBM4 is to completely eliminate the interposers and add memory modules directly to the CPU or GPU matrix. Besides eliminating interposers, the amount of storable memory would increase significantly. This approach is akin to AMD’s 3D V-Cache.
It is likely that SK Hynix and Nvidia will collaboratively design the chip with HBM4, and TSMC may be responsible for manufacturing it using wafer-level bonding technology. However, Nvidia might not be the sole company SK Hynix is partnering with for this endeavor.
Integrating HBM4 alongside the matrix could significantly enhance performance and reduce latencies by eliminating any intermediaries between memory and the logical part of the chip. This, in turn, decreases data processing latency, improving overall performance. Additionally, it would reduce manufacturing costs and simplify chip designs. However, one challenge is the operating temperatures.
“If the heating issue is resolved two or three generations later than now, HBM and the GPU can operate as a single entity without the need for an interposer,” says Kim Jung-ho, a professor in the Department of Electrical and Electronics at KAIST.
“Within 10 years, the ‘rules of the game’ in semiconductors may change, and the distinction between memory and logic semiconductors may become insignificant,” stated a source in the industry to Joongang.co.kr.