Cutting GPU power 42% costs nothing: RTX 3080 Ti runs Qwen 3.5 9B at full speed on 250W
Asthenia5 · reddit · 2026-11-10
A Reddit user benchmarked power-limit scaling on an RTX 3080 Ti running Qwen 3.5 9B Q4KM via llama.cpp:
- Stepping down from 450W in 10% increments, the card held 100 tok/s all the way to 250W — matching stock performance
- Only at 220W did throughput drop (80 tok/s); 260W performed identically to 450W — a 42% power cut at zero performance loss
- GPU utilization stayed high but VRAM bandwidth never maxed out, suggesting the bottleneck lies elsewhere in the inference software or GPU architecture
Takeaway: for small local models, the power limit is far from the performance ceiling, and you can slash power for free.
More from Infra
- vLLM-Omni benchmarks: 2048px image generation in 0.63s on 4x B300 — vllm_project · 2026-11-10
- ReCaVSR one-step streaming 4x video super-resolution ported to ComfyUI — LatentSpacer · 2026-11-10
- Hugging Face livestream today: model quantization with llama.cpp and vLLM — ben_burtenshaw · 2026-11-10
- Running Startlux 4B 9B Instant Q40 on a cluster of 4 Raspberry Pi 5 boards — bartlomiej__ · 2026-11-10
- How banned Nvidia chips actually reach China: stolen trucks, fake paperwork, hair-dryer sticker swaps — Caneki97 · 2026-11-10
- Tencent reportedly picks Bittensor subnet Chutes as launch partner for flagship model Hy4 — markjeffrey · 2026-11-10