July 13, 2026
One person’s trash, another’s GPU farm
Benchmarking 15 "E-Waste" GPUs with Modern Workloads
People are turning “trash” graphics cards into cheap AI beasts — and the comments are obsessed
TLDR: A benchmark project found that old data-center graphics cards can still handle modern AI and rendering jobs for surprisingly little money, making them tempting for home tinkerers. In the comments, people loved the bargain-hunting vibe but immediately argued about power bills, missing hardware, and how far this cheap-card madness should go.
A hacker-style benchmark of 15 retired NVIDIA data-center graphics cards has unleashed exactly the kind of internet energy you’d expect: part budget genius, part electrical safety warning, part full-on dumpster-diving pride parade. The big reveal is that old cards once tossed aside as “e-waste” can still do surprisingly useful modern work for shockingly low prices — think huge memory for under a couple hundred bucks. The author’s pitch is basically: yes, they’re old, yes, they drink power, and yes, that’s exactly why homelab tinkerers love them. In other words, if your hobby involves a loud rack in a garage, this is catnip.
The comment section instantly made it messier and more fun. One camp wanted even more receipts, especially on power draw, with readers basically asking, “Cool story, but how expensive is this space heater really?” Another crowd showed up emotionally attached to obscure bargain chips, with one commenter nostalgically cheering on old PlayStation-linked boards and celebrating that they’re getting a second life instead of rotting as junk. Then came the classic escalation: if these old cards are so cheap, why not stack a mountain of them together and run giant cutting-edge models? That’s where the thread drifted from practical home project into glorious mad scientist fantasy league.
There’s even a sneaky meme-like energy in the replies: one user drops a polished summary “by Mistral AI,” which reads like the comments section outsourcing its own opinion to a chatbot. The mood overall? Delighted, curious, slightly unhinged — and very ready to turn yesterday’s server scrap into tomorrow’s bragging rights.
Key Points
- •The article benchmarks decommissioned NVIDIA enterprise GPUs to assess their usefulness for modern workloads in homelab environments.
- •It cites low used-market prices for older GPUs such as the K80, P100, and V100, positioning them as affordable sources of VRAM.
- •The author pairs old GPUs with inexpensive server hardware including X99 Xeon systems and a Supermicro dual-socket motherboard to explore dense GPU-node builds.
- •The article acknowledges end-of-life limitations such as missing future CUDA and driver updates, but says older software stacks and source builds can preserve usability.
- •The benchmark suite covers computer vision, rendering, Vision Transformer, and large language model workloads, and is published on GitHub.