AI infrastructure is shifting toward a multi-architecture environment, according to a DIGITIMES Intelligence report.
The report noted that competition among cloud service providers is expanding beyond processors to an interconnected infrastructure, enabling accelerators, servers, racks, and clusters to function as a unified system.
“As the number of accelerators and system scale keep growing, performance competition increasingly depends on fabric-based connections among different compute nodes and resources, making interconnect capability a key factor in system scalability and accelerator efficiency,” wrote analyst Joyce Chen.
Multi-architecture environment
The report highlights two major layers of the AI fabric:
- Scale-up: Connects accelerators within a node or rack and prioritises extremely high bandwidth and low latency.
- Scale-out connects multiple nodes, racks, and clusters across larger data-centre environments.
The growing diversity of AI chips creates greater need for different architectures, creating varying requirements for bandwidth, latency, power efficiency, and interoperability, increasing the importance of coordinating these two layers.
Open interconnect standards such as UALink are gaining attention for improving interoperability and reducing dependence on proprietary technologies.
According to the report, the physical technologies used for scale-up and scale-out are also diverging. Copper and electrical signalling remain significant for short-distance, low-latency scale-up connections, while optical interconnects are becoming increasingly important for scale-out to support longer distances, higher bandwidth, and growing power-efficiency requirements.
New opportunities
This landscape shift creates opportunities across several parts of the semiconductor and networking supply chain, including especially for chip vendors, active interconnect IC suppliers, and physical interconnect vendors producing cables, connectors, and optical components.
The shift is particularly significant in Taiwan, where opportunity lies in semiconductor manufacturing as well as switches, optical components, cables, connectors, and system-level manufacturing.
According to the report, AI infrastructure competition will increasingly depend on building faster AI accelerators and scaling diverse computing architectures.
