AMD Stretches Server DRAM with Flash Extended Memory
AMD has introduced a technology that extends server DRAM capacity by using flash storage as an extended memory tier, allowing systems to handle larger datasets without adding costly DRAM. The approach leverages AMD's Infinity Fabric to create a unified memory space, bridging the performance gap between DRAM and NAND flash for memory-intensive workloads.
Background
- AMD is developing "Flash Extended Memory" (FEM), a technology that lets servers use flash storage as if it were additional main memory (DRAM), bridging the gap between fast/expensive DRAM and slow/cheap flash.
- This matters because AI inference and large-scale analytics are increasingly memory-bound — they need more capacity than DRAM alone can economically provide. FEM would let servers handle larger datasets without adding more DRAM sticks.
- The approach builds on AMD's Zen memory controllers, managing DRAM-to-flash translation at the hardware level to minimize software changes.
- This is part of an industry push: Intel had Optane (discontinued), and Samsung/SK hynix are exploring CXL-based memory pooling. AMD's version competes in the same "expand memory without DRAM cost" space.
- Key context: Server DRAM prices remain high, and the rise of large language models (LLMs) means memory capacity often limits what a single server can do. FEM offers a cheaper way to keep more data close to the CPU.