RTX 4070でQwen 35Bを推論すると平均42W — 消費電力プロファイルを4パターン実測Benchmark measurements of Qwen 35B running on an RTX 4070 show average GPU…

Running large language models locally has become increasingly practical on consumer graphics cards, yet questions about real-world power draw and the resulting electricity costs are harder to answer than raw throughput benchmarks. A recent set of measurements published on Qiita addresses this directly, profiling how much power an NVIDIA RTX 4070 actually consumes while running Qwen 35B. The headline figure is notably low, and it offers a concrete reference point for anyone estimating the running costs of a home or small-office inference setup.

According to the report, GPU power during the decode phase — the token-by-token generation stage that dominates most interactive workloads — averaged just 42 W. That sits far below the card's rated Total Graphics Power (TGP) of 200 W. For comparison, the author cites a gaming average of roughly 205 W, meaning that generating text with a 35B-parameter model drew only a fraction of the power the same card pulls during a typical game.

The measurements are broken into four distinct power profiles rather than a single number, which is where the detail becomes useful. While sustained decode averaged 42 W, prompt evaluation — the prefill stage in which the model ingests the input context before generation begins — produced a brief peak of up to 175 W. That still remained below both the 200 W TGP and the gaming figure, but it illustrates that the two phases of inference stress the hardware in very different ways.

The gap between the stages reflects a well-understood characteristic of transformer inference. Prompt evaluation processes many tokens in parallel and is largely compute-bound, so it briefly pushes the GPU's arithmetic units and pulls more power. Decoding, by contrast, produces one token at a time and is generally memory-bandwidth-bound; the compute cores spend much of their time waiting on data movement rather than running flat out. Because power scales with how hard those cores are working, the sustained decode figure appears low even though the GPU is fully occupied with the task.

Some context on the hardware is worth adding. The RTX 4070, built on NVIDIA's Ada Lovelace architecture, ships with 12 GB of VRAM. A 35B-parameter model does not fit into that budget at full precision, so a configuration like this almost certainly relies on quantization — commonly 4-bit formats used by runtimes such as llama.cpp and Ollama — and may offload part of the model to system memory. Partial CPU offloading, if present, would leave the GPU less heavily loaded and could contribute to the low decode wattage, though the report's framing focuses on the measured GPU figures rather than the exact split. Readers reproducing these numbers should note that quantization level, context length, and offloading strategy all influence the outcome.

For cost estimation, the practical takeaway is straightforward. If decode holds near 42 W, an hour of continuous generation consumes on the order of 0.04 kWh from the GPU alone, before accounting for the rest of the system such as the CPU, memory, and power-supply losses. That makes the electricity cost of sustained local inference modest in most regions, and it reframes the common assumption that running a large model is inherently power-hungry. The heavier draw occurs in short prefill bursts rather than across the whole generation, so workloads dominated by long outputs and short prompts are likely to stay closer to the 42 W average.

These figures also sit within a broader industry trend toward efficient local deployment. Open-weight model families such as Alibaba's Qwen, alongside comparable releases, have made mid-sized models accessible on single consumer cards, while inference engines including llama.cpp, Ollama, and vLLM continue to improve memory handling and quantization support. Power profiling of the kind described here complements the more familiar tokens-per-second benchmarks, giving a fuller picture of what a given card can do.

As always, single-configuration measurements should be treated as indicative rather than definitive. Results will vary with driver versions, thermal conditions, the specific quantization used, and the length of prompts and responses. Still, the reported profiles — a 42 W decode average, a 175 W prompt-eval peak, and both figures well under the 200 W TGP and the 205 W gaming baseline — provide a useful, real-world anchor for planning local LLM workloads and their associated energy budgets.