# ADR 027: Retire the legacy NixOS GPU worker

- HTML version: https://robbiepalmer.me/projects/homelab/adrs/027-retire-legacy-nixos-gpu-worker
- Project: Home Lab (https://robbiepalmer.me/projects/homelab.md)
- Status: Accepted
- Date: 2026-10-31
- Supersedes: homelab:007-nixos-gpu-worker

# Context

[ADR 007](/projects/homelab/adrs/007-nixos-gpu-worker) kept an existing Asus
desktop in service as a NixOS batch computer-vision worker and DVC cache. The
machine has an AMD Athlon X4 880K, 16 GB of DDR3 memory, a GTX 1050 with 2 GB of
VRAM, a 120 GB SATA SSD, and a 1 TB mechanical disk. It requires a pinned CUDA
12 and Pascal driver path and currently reaches the tailnet through a tethered
phone rather than a permanent Ethernet connection.

No sustained batch workload emerged after the deployment. The
[Vector workstation](/projects/homelab/adrs/025-graphical-nixos-development-workstation)
now supplies 64 GB of memory, NVMe storage, and an RTX 4080 for the development
and GPU work that the desktop was meant to cover. The lab has no remaining CUDA
12 compatibility requirement.

The tower is too large and noisy for its available locations. Rehousing it
would preserve the old CPU, small GPU, and pinned software while adding case
and cooling costs. Its unused memory capacity does not outweigh the cost of
running and maintaining the complete machine.

# Decision

Retire `asus-desktop` as an active NixOS GPU worker and DVC cache. Do not
rehouse it or preserve its Pascal software path for a hypothetical workload.
Power it only for disk assessment, data migration, and any final inventory
needed for decommissioning.

Keep the current declarative configuration long enough to complete that work.
After no unique data or recovery dependency remains, remove the host from
active deployment, monitoring, and fleet verification. Git history preserves
the old configuration if the implementation ever needs to be studied.

Assess both SATA drives before assigning them another role. The 120 GB SSD is a
candidate Raspberry Pi boot or recovery disk. The 1 TB disk may hold
replaceable data or an additional offline copy only if its SMART report and
self-tests are healthy. Neither disk becomes trusted storage or part of a RAID
array merely because it is available.

Do not retain the motherboard as a speculative lab component. Once the drives
and any required data are removed, sell, donate, or recycle the motherboard,
CPU, DDR3 memory, GTX 1050, case, and power supply as appropriate.

# Alternatives

## Keep it as a 16 GB service companion

This would give the Mac mini more aggregate memory without buying another
computer. It would also make the noisiest, largest, and oldest machine an
always-on service dependency. The hardware has had ample time to find that job
and has not done so.

## Rehouse the components

A smaller case could improve placement. It would not improve CPU performance,
GPU memory, power use, driver support, or storage age. Spending on a case and
cooling would make a later quiet mini computer harder to justify without
solving the Mac's service pressure well.

## Preserve the CUDA 12 worker

The pinned stack still works for narrow Pascal-compatible jobs. There is no
current workload that needs it, while the Vector handles modern GPU work with
far more memory. Unused compatibility creates maintenance without giving the
machine a role.

## Replace it with a mini computer now

A quiet NixOS companion may be useful, but replacement is not part of this
retirement. The separate product proposal must first confirm the Mac bottleneck
and decide whether storage-coupled media services stay together.

# Consequences

The active fleet loses a second local x86 host, 16 GB of memory, and the GTX
1050\. The Vector becomes the only local CUDA machine. No live workload depends
on the retired capacity.

The lab stops carrying a Pascal-specific driver and CUDA maintenance obligation
for an idle computer. Removing the tower also frees the study floor space and
avoids adding its noise to the living room.

Disk reuse remains unresolved until the SMART tests finish. A failing or badly
worn drive will be securely erased where possible and recycled rather than
given a new job. A future companion host will need its own architecture record
and cannot use this retirement as evidence that a purchase is necessary.

---

Markdown index of this site: https://robbiepalmer.me/llms.txt