An experiment with a €49 AMD RX 580: small models and energy use
A €49 graphics card in a 2012–2013 workstation: what small local models can do, and what the energy figures look like. The numbers come from a CLARYEL experiment on 12 November 2026.
On 12 November 2026 an AMD Radeon RX 580 8 GB was installed in P3001, a used workstation with a Xeon E5-2697 v2 processor and 64 GB of DDR3-1333 ECC memory installed in quad-channel mode (the system sees about 62 GB). The card was bought for €49: €45 for the card itself and €4 for shipping. This is the RX 580 2048SP variant: it has 2,048 stream processors, not the 2,304 of a “full” RX 580, and is essentially closer to an RX 570. That distinction matters when comparing it with other cards in the same line. The operating system is Ubuntu 24.04.
Strata will not run on this card, and we do not plan to ask for support of this GPU model in Strata: the author’s time is better spent on more capable hardware. The experiment used llama.cpp with the Vulkan backend. There is no CUDA, no tensor cores and no FP16 acceleration. On the same workstation, a small 3B model at Q4 quantization reached 60 tok/s in decode. For comparison, on a large 77.9 GiB Qwen model, RX 580 with llama.cpp measured 33.9 tok/s prefill and 5.85 tok/s decode. The same workstation with an RTX 3050 8 GB and Strata, on this workload, measured 128–188 tok/s prefill and 23–31 tok/s decode. During inference the RX 580 drew up to about 120 W. For a model of that size the card is the weaker option; as a “second brain” for short micro-tasks, the results still make it useful, and it costs less than a comparable amount of newer system memory.
The experiment then moved away from the largest possible model and toward small models that fit entirely in 8 GB of VRAM. Three models were measured separately: Qwen2.5-3B Q4 — 1.79 GiB, 392 tok/s prefill and 60 tok/s decode; Nemotron Nano 4B Q4 — 2.63 GiB, 268 / 40 tok/s; and Bonsai 2 9B Q2 — 2.37 GiB, 170 / 27 tok/s. Combined weights come to 6.79 GiB. In theory all three could fit in 8 GB, leaving roughly 1.21 GiB for KV cache, compute buffers and overhead. That is tight, and simultaneous residency of all three in VRAM has not been verified.
GPU power draw during local inference was also measured at five power limits: 120, 100, 70, 50 and 30 W. For small models that fit entirely in VRAM, 70 W was the most efficient point for generation in this experiment. Compared with 120 W, GPU power during inference was about 36% lower, generation speed only 12–15% lower, and energy per generated token 26–28% lower. At 70 W the 3B model generated 39.5 tok/s. A short 200-token answer used 0.091 Wh of GPU electricity. At €0.25/kWh that is €0.000023 per such an answer, €0.023 for a thousand answers a day, and about €8.30 for 365,000 such answers in a year. This estimate excludes idle power and the rest of the computer. The RX 580 draws about 14 W when idle, equivalent to about €30.70 per year if left powered continuously; the CPU, memory, motherboard and storage add more. For low-volume workloads, cutting idle power can matter more than cutting inference power. A caveat: 70 W is optimal for the small-model generation workload measured in this experiment, not necessarily for every task. Long prompts, OCR pipelines, audio processing and other workloads may have entirely different speed and energy characteristics.
The next line of work is not “how large a model can be run”, but how many useful everyday tasks an inexpensive local system can perform: parsing small PDFs, extracting data from scans, recognizing text in pictures, transcribing short recordings, classifying files, summarizing documents, reading logs, routing requests and preparing data for a larger AI system. Those multimodal and document-processing workloads have not yet been experimented with on this card. Their feasibility, accuracy, latency, memory need and energy cost all still have to be investigated. Some operations may be better handled on the CPU, leaving the GPU for inference. Whether several specialized models can remain loaded, with only one running at a time, is also still an open question.
Power delivery is a separate engineering question. A standard PCIe x16 slot can supply up to 75 W directly from the motherboard. The most efficient mode for small models in this experiment was a 70 W software limit. The idea under consideration is a configuration that runs from the slot alone, without an extra 6-pin or 8-pin power connector, so that such a card could be fitted in an existing office computer without a power-supply upgrade. That is a direction of work, not a finished solution. In the experiment, even with the 70 W software limit, brief peaks above 75 W were still observed, and the RX 580 we have still needs auxiliary power. A safe slot-only design would need engineering work on peak power and compatibility checks. Document, image and audio workflows also still need to be confirmed. We are not stating that such an installation is already being supplied to customers.
A €49 card does not replace the main Strata system. On the results already in hand, it can become a low-cost, relatively economical worker for many short, repetitive tasks, while keeping data local. All measurements, power-limit experiments, method, raw logs, failures and conclusions are published in the repository. We are exploring how much useful local AI can be obtained from hardware that is often treated as obsolete.