HBM-connected GPUs shifting toward optical interconnects - 富途牛牛

As artificial intelligence workloads continue to expand, demand for memory capacity and bandwidth is pushing existing GPU-HBM architectures toward their limits. Researchers at a major South Korean mem... Read original →

The AI compute arms race is forcing a decoupling of the physical GPU-HBM bond. Separating them into distinct packages linked by optical interconnects appears as an engineering workaround for 20+ HBM layer yield limits—but actually signals a paradigm shift from 'tight integration' to 'functional modularity.' This redefines OSAT value chains: TSMC’s (Taiwan, China) dominant 2.5D CoWoS faces new barriers from photonic integration, opening doors for Intel and Ayar Labs. Geopolitically, reliance on III-V materials and lasers—subject to U.S. export controls—means NVIDIA’s adoption could trigger stricter BIS scrutiny for Chinese customers. Within 18 months, standardization of optical I/O at the system level will become a battleground, while server form factors adapt to fiber routing. This isn’t incremental—it’s an architectural rupture in AI infrastructure.