How I built a custom cooler to tame retired server hardware for modern AI and compute tasks at home.

There’s something incredibly satisfying about taking old, discarded tech and giving it a new purpose. My latest project has been a deep dive into exactly that, focused on a fascinating challenge: how to repurpose enterprise GPUs for modern, heavy-duty tasks right here in my home lab. I’m talking about those powerhouse NVIDIA Tesla cards that were once the heart of data centers. They’re amazingly powerful for their age, but there’s one big catch—they were never designed to cool themselves.

Enterprise cards are built to live in server racks with jet-engine-level fans forcing massive amounts of air over them 24/7. Take them out of that environment, and they’ll overheat in minutes. So, my goal became creating an “Endgame EOL’d GPU Box” — a machine that could put this cast-off gear to work on modern AI and compute loads. And the biggest hurdle? Building a totally custom air cooler from scratch.

The Challenge: Why You Can’t Just Repurpose Enterprise GPUs

When you get your hands on a card like a Tesla M40 or P100, the first thing you notice is the lack of fans. It’s just a massive heatsink. My first few attempts at cooling were, let’s say, learning experiences. It quickly became clear that a standard off-the-shelf cooler wasn’t going to cut it. I had to design and build my own.

After several iterations, I’ve finally landed on a prototype cooler that I think can handle the thermal load. It’s not the prettiest thing in the world, but it’s built for one thing: performance. Now that the hardware puzzle is mostly solved, the real fun begins: benchmarking. How does this old-school hardware stack up against modern demands?

The Hardware Gauntlet: My Benchmarking Plan

To see what this setup is really capable of, I’ve put together a collection of GPUs to test with my new cooler. The idea is to see how performance scales, not just with single cards, but with multiple GPUs working together.

My GPU Lineup Includes:

• NVIDIA Tesla K80s (an older, interesting dual-GPU card)
• A mix of the “Maxwell” generation: M10, M40, and M60
• The “Pascal” powerhouses: the P40 and P100
• A modern beast for comparison: the “Volta” V100

I’ll be running tests with single cards, and also in multi-GPU configurations, like four P100s at once. One thing to keep in mind here is PCIe bandwidth. Cramming that many GPUs into a system can force the motherboard to downgrade connection speeds to some of the cards, which can create bottlenecks. Part of the test is to see how much that really matters. For more on how that works, AnandTech has a great explainer on PCIe lanes.

To make sure the GPUs are the star of the show, I’ll be running these tests on two different, but very capable, Intel Xeon CPUs (an E5-2687W v4 and an E5-1680 v4) to see how the processor choice influences the results.

The Software Gauntlet: Real-World Tests for these Repurposed Enterprise GPUs

A pile of hardware is useless without the right software to push it. I want to test these cards on tasks that people are actually running today, especially in the AI and scientific computing worlds.

My testing playbook will run a series of benchmarks automatically:

1. AI Language Models: I’ll be testing throughput using vLLM, a popular library for running large language models, loaded with the Llama 3 8B model. This is a perfect test of modern AI inference performance.
2. Scientific Computing: I’ll have the GPUs contribute to distributed computing projects like Folding@Home. It’s a great way to test sustained load while also contributing to scientific research. I’ll also throw some other BIONIC and Einstein@Home workloads at them.
3. General AI Benchmarks: I’ll run a few standard suites like ai-benchmark.com and llama-bench to get some standardized numbers that are easy to compare.

This project has been a ton of fun, blending hardware fabrication with deep-dive software testing. It’s a quest to see if I can give this powerful, forgotten hardware a second life. The next step is to run the gauntlet and see what the numbers say. I’ll be sure to share the results once they’re in. It’s amazing what you can build with a little curiosity and a pile of old server parts.