Build note
ThinkStation P920 Server
The machine is a Lenovo ThinkStation P920. It is a workstation with two Xeon CPUs, 124 GiB of RAM, and two NVIDIA GPUs. The machine is the first decision in the series, and it is the one that shapes everything that follows. The other decisions are the GPU split, the disk split, and the OS choice.
What is in the box
| Part | What it is |
|---|---|
| Machine | Lenovo ThinkStation P920, 30BDS0RA00 |
| CPUs | Two Xeon Gold 6148, 80 threads total |
| Memory | 124 GiB |
| GPU, compute | RTX 3090, 24 GiB, at PCI 3b:00.0 |
| GPU, display | RTX 3060 LHR, 12 GiB, at PCI af:00.0 |
| OS | Ubuntu 24.04.4 LTS |
| Driver | 595.99.02, CUDA 12.8.93 |
The dual-socket board is the part that matters. A desktop has one CPU and one memory bus; this has two, and the memory is split between them. A thread that runs on the wrong socket reads the other socket’s memory over the interconnect, which is slower than the local memory. That is a footnote for a desktop and a running cost for a server, because the server is a lot of threads.
Three decisions, and the rest of the series follows them
The GPU split. The 3090 is the compute card and the 3060 is the display card. The 3090 stays headless, so the compositor and the model never compete for the same memory. The 3060 has the connectors, so the monitor stays on the desktop while the model runs. Part 2 covers the split, and the number that makes it a trap: the same two cards have three indices.
The disk split. The home directory, the data and the snapshots are separate filesystems. The system is a separate filesystem, and the snapshots do not cover it. That means a rollback is a system-only rollback, and a model that is not data does not get a free copy. The 833 GiB NTFS disk is a separate case: it is not mounted, and the series does not pretend it is.
The OS is the default. Ubuntu 24.04 LTS, not a rolling release, not a container for the host. The driver is the thing to keep, and the driver is the thing that the kernel, the Secure Boot setting and the kernel headers all have to agree with. Part 3 is the kernel and DKMS story.
What the server is not
The server is not a server in the sense of a box in a rack. It is the workstation with the desktop, and the desktop is the reason the GPU split is the way it is. The server is not a rack, so “no snapshots on root” is a deliberate choice, not a missing piece.
The rule that would have prevented the next week
Keep the machine’s inventory in one file. The card, the driver, the index, the kernel, the port, the disk and the file all change, and the next time one of them changes, the file is the only thing that says what the others were. The inventory is the source for every part after this one, and this part is where it starts.
How this was made
- Written by: Rupesh, from his own
~/system-inventory.mdand server notes, with AI drafting and editing. - Numbers: from
nvidia-smi,/etc/os-releaseand the inventory note. - AI-assisted sections: structure and copy editing.