Meta and Panmnesia are developing a new data center architecture called CXL, which aims to connect up to 960 AI accelerators—essentially powerful computing chips—into a single, unified system across multiple server racks. This design uses a technology called Compute Express Link (CXL), which allows CPUs, accelerators, and memory to communicate directly with each other without relying on traditional network connections like Ethernet or InfiniBand. This could enable nearly 1,000 GPUs to function as one cohesive computing system, improving efficiency and performance in large-scale AI training. One of the main challenges in AI training is coordinating hundreds or even thousands of accelerators that process massive amounts of data. These accelerators must go through repeated computational steps, and if one device is delayed, it can slow down the entire system. The proposed CXL architecture focuses on reducing unpredictable delays that occur when data moves between server racks. By using CXL, the system can maintain a consistent communication flow, which is crucial for maintaining performance in high-speed computing environments. The architecture includes specialized hardware such as a high-fan-out switch, a link acceleration unit, and a fabric controller to manage communication across the system. These components are organized into modular units called trays and pods, similar to how blocks are arranged on a semiconductor chip. Panmnesia has already completed silicon validation for its fabric controller and link acceleration unit, and the switch has been fabricated. Some pre-release silicon is currently being used to refine the design as the company moves toward commercial products. Compared to existing systems like NVIDIA's GB200 NVL72, which connects one CPU to two accelerators, Panmnesia's architecture allows a single CPU to coordinate up to 16 accelerators—a significant improvement. This scalability means that about 60 such groups could be combined into one coherence domain, housing around 960 accelerators. The system also promises faster cross-rack communication, potentially reducing delays from microseconds to hundreds of nanoseconds. Additionally, it allows for individual device replacement without shutting down an entire server, potentially improving system reliability and reducing hardware waste. Panmnesia CEO Myoungsoo Jung emphasized that CXL could enable an entire data center to function as a single computing unit. However, current CXL technology using electrical signals is limited to about seven meters at high speeds. To overcome this, Panmnesia is exploring optical CXL links for longer distances and has already validated this approach in hardware. This innovation could be a key step toward building more powerful and efficient AI computing systems in the future.