Google Reuses Retired Server RAM in New AI Servers Amid DRAM Shortages

Google is turning to creative solutions to address the ongoing memory constraints in the DRAM sphere. According to Nikhil Cherian, senior director of Google's supply chain infrastructure, the AI industry has shifted from being compute-constrained to memory-constrained, with high-performance memory accounting for around 75% of an AI server's bill of materials costs.

To 'break through the memory bottleneck,' Google is working on both software and hardware solutions. In a bid to optimize its resources, the tech giant has established an internal recycling supply chain by reusing DDR4 modules from retired servers in new AI servers. Cherian revealed that Google has even designed special hardware adapters to connect previous-gen memory solutions like DDR4 to the latest generation of its AI servers.

This move is part of a broader effort to reduce costs and improve efficiency. Google's TPUs (Tensor Processing Units) are optimized for AI workloads, with the TPU8i model leveraging highly advanced, specialized memory layers to eliminate performance bottlenecks. However, in some instances, Google appears to be reusing older DDR4 modules in place of newer DDR5 modules within its new AI servers.