My Homelab - Hardware
Overview
My infrastructure is composed of three machines: a development workstation and two Proxmox nodes that form a cluster.
Given the hardware shortage, it's necessary to think carefully about intelligently distributing resources between machines, to use them efficiently. Here's how I thought it made sense to organize resources:
Comparison Table
| Dev PC | PVE | PVE-2 | |
|---|---|---|---|
| Role | Development workstation | Main Proxmox node | Secondary Proxmox node |
| CPU | AMD Ryzen 7 5700G (8c/16t, 3.8 GHz) | AMD Ryzen 9 9950X3D (16c/32t, 5.8 GHz) | Intel Core i7-10700K (8c/16t, 5.1 GHz) |
| RAM | 32 GB DDR4 | 64 GB DDR5 | 32 GB DDR4 |
| GPU | RTX 3060 Ti | RTX 5090 + RTX 5060 Ti | RTX 3090 |
| Disk 1 | Samsung 970 EVO Plus 2 TB (NVMe) | Samsung 9100 PRO 4 TB (NVMe) | Kioxia 1 TB (NVMe) |
| Disk 2 | — | — | Seagate 8 TB (HDD SATA) |
| Disk 3 | — | — | Toshiba 4 TB (HDD SATA) |
| Disk 4 | — | — | WD 8 TB (HDD SATA) |
The Dev PC:
Its purpose is to run all my projects at the same time. (PyCharm, Webstorm, Firefox, my SSH client...). For that, 32GB of RAM is sufficient, although often a bit tight. It has 2TB of disk and that's enough: the datasets / data I use in my projects are stored on PVE-2.
PVE:
With two Blackwell GPUs, a Ryzen 9 9950X3D, 64GB of DDR5 and 6TB of NVMe (2TB NVMe4 and 4TB NVMe5), this is the flagship of the homelab. It hosts my entire AI stack (LLMs, embedding models for RAG, ASR / TTS for my voice assistant...) and it's with this that I work on all my Machine Learning experiments.
PVE-2:
This is the original node, which holds most of the storage and on which I host my services. (Gitlab, Pi-hole, Garage S3, media services...)
The GPUs
... It's important in our time, right?
As I sometimes tell my friends, "in life there are people who like fine mechanics, seeing beautiful cars, hearing engines roar. Well, what I love is graphics cards".
My first graphics card dates back to 2021, I managed the small feat of buying an RTX 3090 Founders Edition at MSRP (Nvidia's defined price) right in the middle of the Covid shortage. During a drop, cards are available for about 30 seconds on LDLC's website, and you need to be fast to get one. For my RTX 3090, the site showed an error when adding to cart, even though the site still showed "available". I kept clicking, and after 20 minutes of chaining errors I managed to buy it... So, patience and perseverance, they pay off.
My older brother asked me at the time "but what are you going to do with 24GB of VRAM? That's huge", to which I replied "don't worry it's for training models". With DealTracker, bet paid off.
My 3060 Ti, I bought it after the Covid crisis, I found a pretty nice PC on LeBonCoin (€500 for a prebuilt Ryzen 7, 16GB DDR4 and a 3060 Ti! Can you believe it?)
And ultimately, I went from 2 GPUs to 4 recently (early 2026), as I watched hardware prices explode. Because at a time when Nvidia is abandoning consumers in favor of datacenters, it seems sensible to secure access to your own computing resources. It's the key to digital sovereignty, free AI usage, and Machine Learning projects with controlled costs.
So I bought a prebuilt with RTX 5090, and added an RTX 5060 Ti 16GB in the same case. It works power-wise because I undervolted the 5090 to 450W (given all the cable melting issues we've seen on Reddit...). Here I come for the benefits of the Blackwell architecture!
My UPS
After investing in good hardware, it seems natural to want to protect it, to avoid any trouble.
"What, the RTX 5090 doesn't work anymore after the power outage? That's too bad, there must have been a voltage spike when they restored power... I hope you're ready to send emails to the seller to explain that you want to make a warranty claim on a product costing several thousand euros and wait several months..." <== I don't want that to happen to me
The most reliable UPS units can cost quite a bit, but they're robust and last for years. It seemed very sensible to buy one second-hand, by buying a good reference in good condition I ensure I have solid equipment, and by replacing the batteries I start fresh. After research, I started looking for a Smart-UPS 1500 on LeBonCoin. The idea was to spend ~€300 to get a UPS worth ~€900 new. I found a listing that looked good, not too far from my place. I went to the meetup with a suitcase, this UPS weighs 24kg after all...
Arriving on site, I talk to the seller, who was very nice. The UPS is indeed very very heavy... And there on the front, I see it says "Smart-UPS 3000". Confused, I sent a quick message to an LLM to ask its opinion, barely having time to see the beginning of its response.
"Well now, we're not AT ALL in the same category of UPS as the Smart-UPS 1500...".
I smiled at the seller and told him "I'll take it". So even without DealTracker, I have a certain talent for stumbling across sellers who don't know what they're selling.
On the way back home, I look at the details. The Smart-UPS SMT3000IC costs €2,000 new, and its max power is 2700W... Perfect. Just need to bring this 53kg monster home, change the batteries, and I'll have datacenter-worthy protection for my equipment.
Installation and Network Setup
Having two Proxmox nodes is nice, but if they can talk directly to each other, that's better. It reduces inter-VM latency, smooths out heavy data transfers in case of a VM migration from one node to another...
Except problem, consumer motherboards are not designed for this at all. On professional models there are SFP+ ports, but on my two nodes... not at all, I just have a single 2.5Gbps RJ45 port. My first node, I took a micro-ATX board with the idea of recycling it as a NAS later, and the second node I took a prebuilt because of the hardware shortage.
So what do I do, add a network card that gives me more RJ45 / SFP+? Ah no, that's going to take up PCI-e lanes, and I simply don't have the physical space for that... So maybe I need two USB-C-RJ45 adapters? And I connect the two PCs that way? Yeah that could work!
It's at this point that Malrick adds his RTX 5060 Ti to the second node.
"Incredible, I can't believe it. Adding a second GPU disabled the native RJ45 port. I see nothing in the BIOS that allows fixing this..."
Ultimately, rather than fight with consumer hardware making compromises on PCI-e routing, I decided to give up the idea of connecting them directly. For my use, it doesn't seem necessary at all. Often, the perfect is the enemy of the good.