Status: In Progress

Scope:

Building a custom multi-GPU server from server-grade components, both as a hands-on hardware and infrastructure project and as a long-term platform for self-hosted AI work — running and experimenting with my own AI agents, assistants, and LLMs once operational.

Process:

Selected server-grade hardware built around a Supermicro H12SSL motherboard, an AMD EPYC 7443P (Milan) processor, and 256GB of DDR4 ECC memory. Sourced and installed four NVIDIA RTX 4070 Ti Super GPUs — since the cards didn’t fit inside the original chassis, designed an open-air mounting configuration above the main unit using PCIe risers, with generous spacing between cards for airflow. Wired dual power supplies with synchronized startup, splitting the load across both to handle sustained power draw well over a kilowatt under full load. Built out a cooling design combining the open-air GPU layout, supplemental intake fans, and rack-level exhaust with an external temperature probe for monitoring.

Debugging:

Initial power-on attempt failed to POST. Diagnosed the issue as incorrect motherboard power cabling rather than a hardware fault — corrected the cable configuration and achieved a successful POST.

Progress So Far:

Installed Ubuntu Server 24.04 LTS. Verified system integrity post-install: confirmed all four GPUs correctly detected via lspci, confirmed storage partitioning and mount points via lsblk and df, and validated system information via dmidecode. Completed cable management and finalized the rack-mounted hardware layout.

Next Steps:

Connect the system to the network to complete remaining setup, including NVIDIA driver installation and four-GPU compute validation. Extend the root logical volume to use the full available disk capacity before deploying any storage-heavy workloads. Proceed to IPMI/BMC remote management configuration and thermal validation under real sustained load.