Today I had another unscheduled event–I had to physically recreate my homelab completely from scratch, which meant unplugging all my machines, completely emptying the closet they live in, and then painstakingly re-assemble and re-wire the whole thing in haste.
I had originally intended to spend some quality time testing an ONNX model I intended to quantize and test on ARM devices (I need something small but efficient to do embeddings, and will post about that soon), but this popped up instead.
The reason why this happened is that my UPS started failing again after 5 years and I decided to get a new one, which arrived while I was in hospital.
I put it aside while I recovered, but today the battery alert fired again and I decided it was time to do the swap–only to find that the new UPS didn’t fit the spot I had for it, so I had to reshuffle the whole closet and cobble up a temporary solution:

The current setup is definitely improvised and the cable routing looks really bad, but I am already designing a “mini-rack” type setup that I will 3D print to hold all the power bricks and route the cables properly, as well as a couple of braces to hold the UPS a few centimeters off the floor–because for some reason Schneider/APC still has non-removable power cords in these models and doesn’t provide anything suitable to raise the UPS just enough off the floor to prevent water leak damage.
The old model was rated for 950VA and the new one is rated for 900VA, so there isn’t that much of a difference. We only rarely have power fluctuations that at most translate to short power outages (other than the recent blackout), so the difference shouldn’t really matter.
The real improvement (for me) is that the new model has a small display that clearly shows status (logging in to a server to have it check the UPS at 2AM is not something I want to keep doing), and, most importantly, a mute button for alerts.

The UPS is roughly at 40% load, and that is actually because I have more machines in my closet now–I took advantage of the reshuffling and finally moved borg
there after almost two years trying to sort out its thermals.
I also shuffled things around to see if I could improve thermals (by moving most critical machines closer to the ground and emptying the closet of clutter that impeded airflow), and I’m sort of happy with the results for borg
for now but not at all happy with the am18
, which idles at 52oC but peaks at 89oC running Bazzite.
I have been meaning to replace that machine’s underside fan and 3D print a little riser to improve airflow, and that is definitely happening as well.
Right now the closet temperature is stable at 31oC with the doors shut (for comparison, it’s 27oC ambient in my office), and that is almost four degrees below what it was before I did this.
So even though this was completely unplanned and took a huge chunk of my free time I am not getting back this weekend, I’m rating it a win. I have plans to improve air circulation, but those will have to wait until the basics are sorted out.