Conceptual illustration of decommissioned GPUs reborn in a homelab server rack
E-waste GPUs can still run useful modern workloads, but only the right old cards are worth buying. A 2026 benchmark of 15 decommissioned enterprise GPUs found that some reseller Tesla cards still deliver real LLM, vision, and transcription performance for a few dollars, while others are barely useful for experiments. The winning models depended more on workload than raw age.
I read the primary benchmark directly on esologic.com; cross-checked context and reported prices on the live Hacker News thread, then confirmed publication traction in blog.donweb.com’s Spanish summary using Jina-based verification of those pages. The prices and standout results below reflect that cross-checked source set.
For builders choosing between a used GPU box and a lean main workstation, a queued slow-social productivity guide offers a useful analogy: choose hardware by the friction you actually need, not by maximum badge value.
Featured image: C:\Users\hbloo\tekmag-e-waste-gpu.jpg — generated via FLUX.1-schnell; upload/attachment handling is deferred to Dev.
What They Actually Benchmarked
The test suite used a Docker-based 17-test pipeline, which matters because reused enterprise cards can score well on raw throughput yet still fail on prompt-heavy models or warm real-world workflows. Workloads included LLM inference, vision inference and training, vision transformers, Whisper-style audio transcription, storage-to-GPU bandwidth, and rendering and compute baselines.
Hardware covered Tesla K80, M10, M40, M60, P40, P100, V100-16GB, and T40, each measured against buying prices reported on reseller markets. The most interesting price clusters were K80 cards around $30 to $60, P100 near $50 to $150, and V100-16GB near or under $200. Several cards from Kepler and Pascal eras were cheap enough to experiment with, yet their value was narrowly workload limited.
Standout Findings That Matter for Buyers
One of the stronger market signals was that the P40 outperformed the P100 on LLM workloads despite being an earlier architecture. That kind of workload reversal is why shopping by generation alone often fails with e-waste hardware. On the community side, Hacker News commenters noted Tesla P4 setups achieving steady decoded throughput around 7 to 12 tokens/sec on 20B to 30B parameter Q4 quantized models, with the main limitation being prompt loading rather than continuous decoding.
The M60 stood out for audio transcription, while the V100-16GB tracked close to T40 performance across broader workloads. That combination gives buyers a clearer shopping rule than “buy the youngest-old card”: match the card to the actual task profile, then buy the cheapest acceptable unit for that workload.
Build Constraints Worth Honoring
The good news for homelab builders is that thermal fit is often easier than expected. With aftermarket coolers, Tesla-style cards can fit densely in standard ATX cases and still leave room for a 10GB NIC. That makes datacenter e-waste physically practical even in smaller builds.
The harder part is software. Older cards often require older CUDA stacks, and some Kepler-era hardware only worked in this benchmark because the author used custom Docker builds. If you care about long-term maintainability, choose hardware whose tested stack still compiles and runs on a current host OS. For builders seeking cleaner nearby software choices, see our app curation guide for reducing environment friction around low-power boxes.
When E-Waste GPUs Are Actually Worth It
For budget inference experiments, audio transcription, learner compute, or visualization practice, an old enterprise card can be a smart purchase and the reuse argument is easy to defend. For 24/7 production inference, models with worn fans and unclear thermal history are harder to justify. In that case, a cleaner software environment is often a better return on money and power.
My take: this benchmark is a workload reference, not a flat value ranking; reseller prices vary too much for a single winner. Start with the task you want to run, then buy the cheapest tested card that satisfies it.
How to Choose
- Match workload first: choose the card that the benchmark proved strongest for your specific task.
- Thermal condition: inspect fan and VRAM temperatures before buying retired datacenter cards.
- Software support: verify CUDA, driver, and host OS compatibility before committing.
- Power and fit: check PSU headroom, PCIe generation, and physical case clearance.
- Price discipline: compare completed eBay sales, not only active listings.
Frequently Asked Questions
Are old GPUs still useful for AI tasks?
Some are. Newer retired datacenter cards can still run local inference and transcription, while older cards may need dated software stacks or only work for casual experiments.
Which old GPU should I buy first?
It depends on the workload. For LLM inference the P40 and P4 are interesting; for broader tasks the V100-16GB is closer to newer T40 performance and may be worth the extra $50 to $150.
Should I run old GPUs 24/7?
Usually not. Used datacenter cards can have aged fans and thermal history; inspect them carefully before using them for always-on inference or training workloads.
References
- esologic — Benchmarking 15 “E-Waste” GPUs With Modern Workloads
- Hacker News — Benchmarking 15 E-Waste GPUs discussion
- blog.donweb — GPUs e-waste para IA: qué rinden de verdad en 2026
Originally published on TekMag
Top comments (0)