Old server machine with 2x 3060 12GB for MoE models, anything I’m missing?

Explain Like I'm Normal

A community member is repurposing a legacy HP Z640 server with dual Xeon processors and adding dual RTX 3060 GPUs to run Mixture of Experts (MoE) models. This setup highlights the growing trend of 'AI scavenging,' where builders use high-VRAM consumer cards and server-grade ECC RAM to bypass expensive enterprise hardware. The main challenge remains managing PCIe lanes and NUMA nodes across dual-socket CPUs.