NVIDIA Unveils CDMM Mode to Optimize GPU Memory Management on Coherent Platforms
NVIDIA has rolled out Coherent Driver-based Memory Management (CDMM), a new mode designed to enhance GPU memory control on hardware-coherent platforms like GH200, GB200, and GB300. The innovation targets performance inconsistencies caused by non-uniform memory access (NUMA) in systems where applications lack full NUMA awareness.
Under NUMA mode—the current default—GPU memory is exposed to the operating system as a generic pool, allowing dynamic migration between CPU and GPU but risking suboptimal performance. CDMM shifts this paradigm by letting the Nvidia driver manage GPU memory directly, mirroring the efficiency of PCIe-attached GPU models. The change promises finer control and potential performance gains for developers and cluster administrators.
