CPU performance improvements based on Dennard scaling and Moore's Law have already reached their limits, and domain-specific computing has been considered as an alternative to overcome the limitations of traditional CPU-centric computing models. Domain-specific computing, seen in early graphics cards and network cards, has expanded into the field of accelerators such as GPGPUs, TPUs, and FPGAs as machine learning and blockchain technologies become more common.

The exponential growth in data volume and traffic across the global internet infrastructure is motivating the exploration of new architectures for data centers to achieve the demand for better performance, efficiency and total cost of ownership (TCO). One promising trend is to segregate the functional components of compute, memory, storage, and networking into pools and organize them as needed according to the specific needs of different workloads. This disaggregation and composability will take us beyond the classic architecture of servers as the unit of computing.

CXL has garnered increasing attention from the industry, with a predominant focus on memory expanders utilizing DRAM-based technology. However, the discontinuation of Optane and the growing demand for extensive memory in new ML models have prompted the storage and system communities to urgently seek alternative solutions. In a keynote address last year, Dr. Ki presented compelling reasons for the storage community to seriously consider CXL adoption.