Research Papers 论文研究 5h ago Updated 2h ago 更新于 2小时前 45

Energy Efficiency of Locally Deployed LLMs: A Preliminary Quantitative GPU Power Benchmark on Consumer Hardware 本地部署LLM的能源效率:消费级硬件上的初步定量GPU功耗基准测试

First reproducible hardware-level energy benchmark measuring GPU power draw of nine open-source LLMs (1B–7B parameters) on consumer RTX 4060Ti 16GB hardware using Ollama gemma3:1b and llama3.2:1b achieve the lowest energy cost at 0.56 J/token and 0.65 J/token respectively, with throughput exceeding 170 tok/s 7B-Mistral consumes up to 4.4x more energy per token than the most efficient model, demonstrating that parameter count alone is a poor predictor of energy efficiency qwen3.5:2b shows anomalo 在消费级GPU(RTX 4060Ti 16GB)上对9个1B-7B参数开源LLM进行硬件级能源效率基准测试,填补本地部署能耗研究空白 gemma3:1b和llama3.2:1b能效最优,分别仅需0.56 J/token和0.65 J/token,吞吐量超170 tok/s 7B-Mistral每token能耗高达高效模型的4.4倍,证明参数量并非能效唯一决定因素 模型架构与量化策略对能源效率影响显著,qwen3.5:2b因扩展推理模式导致能耗异常 研究提出区分token生成模式的能效评估方法,为本地LLM部署提供可复现的基准框架

58
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • First reproducible hardware-level energy benchmark measuring GPU power draw of nine open-source LLMs (1B–7B parameters) on consumer RTX 4060Ti 16GB hardware using Ollama
  • gemma3:1b and llama3.2:1b achieve the lowest energy cost at 0.56 J/token and 0.65 J/token respectively, with throughput exceeding 170 tok/s
  • 7B-Mistral consumes up to 4.4x more energy per token than the most efficient model, demonstrating that parameter count alone is a poor predictor of energy efficiency
  • qwen3.5:2b shows anomalously high per-prompt energy due to extended internal reasoning, revealing a critical gap in current efficiency metrics that fail to distinguish between token generation modes
  • Model architecture and quantization strategy are identified as primary drivers of energy efficiency, not just raw parameter count

Why It Matters

As local LLM deployment grows for privacy and on-premise inference, practitioners need concrete data on energy costs to make informed hardware and model selection decisions—yet the field has overwhelmingly focused on accuracy benchmarks while neglecting power consumption. This paper fills that gap by providing reproducible, hardware-level measurements that directly inform the total cost of ownership for edge and consumer-grade AI deployments.

Technical Details

  • Benchmark setup: Nine open-source LLMs ranging from 1B to 7B parameters evaluated on a single NVIDIA RTX 4060Ti 16GB GPU using the Ollama inference engine, with GPU power draw sampled at 2Hz via nvidia-smi across a fixed prompt set
  • Metrics measured: Mean and peak power consumption, total energy per prompt (J/prompt), energy per output token (J/token), and throughput (tok/s)—providing a multi-dimensional view of efficiency beyond single-metric benchmarks
  • Key findings: gemma3:1b (0.56 J/token) and llama3.2:1b (0.65 J/token) led in energy efficiency with >170 tok/s throughput; 7B-Mistral consumed up to 4.4x more energy per token than gemma3:1b despite only ~7x more parameters, confirming architecture and quantization dominate over parameter count
  • Anomaly detection: qwen3.5:2b exhibited unexpectedly high per-prompt energy attributed to extended internal reasoning processes, highlighting that current benchmarking methodologies do not adequately account for different token generation modes (e.g., chain-of-thought vs. direct response)

Industry Insight

  • Organizations deploying LLMs on consumer or edge hardware should prioritize architecture and quantization choices over raw model size when optimizing for energy efficiency; a 1B-parameter model can outperform a 7B model on both energy and throughput depending on design choices
  • Benchmarking frameworks and evaluation pipelines must incorporate energy-per-token metrics alongside accuracy and latency to reflect real-world operational costs, especially as sustainability becomes a regulatory and business priority
  • The qwen3.5:2b anomaly suggests that models with extended reasoning capabilities may appear inefficient under standard benchmarks, indicating a need for mode-aware evaluation protocols that distinguish between reasoning-intensive and direct-generation inference patterns

TL;DR

  • 在消费级GPU(RTX 4060Ti 16GB)上对9个1B-7B参数开源LLM进行硬件级能源效率基准测试,填补本地部署能耗研究空白
  • gemma3:1b和llama3.2:1b能效最优,分别仅需0.56 J/token和0.65 J/token,吞吐量超170 tok/s
  • 7B-Mistral每token能耗高达高效模型的4.4倍,证明参数量并非能效唯一决定因素
  • 模型架构与量化策略对能源效率影响显著,qwen3.5:2b因扩展推理模式导致能耗异常
  • 研究提出区分token生成模式的能效评估方法,为本地LLM部署提供可复现的基准框架

为什么值得看

随着隐私需求推动本地LLM部署,能耗成本成为关键考量,但消费级硬件上的量化研究仍显不足。本文提供了可复现的硬件级基准测试方法,帮助从业者在精度与能效间做出更明智的选型决策。

技术解析

  • 测试平台与数据采集:单卡RTX 4060Ti 16GB,使用Ollama推理引擎,通过nvidia-smi以2Hz频率采样GPU功耗,覆盖平均/峰值功耗、每提示总能耗(J/prompt)、每输出token能耗(J/token)、吞吐量(tok/s)
  • 评估模型集:9个开源LLM(1B-7B参数),涵盖gemma3、llama3.2、Mistral、Qwen等主流架构,采用固定提示集进行对比
  • 核心发现:1B参数模型能效显著优于7B,gemma3:1b以0.56 J/token领先;7B-Mistral能耗达4.4倍;qwen3.5:2b因扩展推理模式出现能耗异常
  • 方法论创新:区分普通token生成与扩展内部推理模式,提出更精细的能效评估框架,避免单一指标误导

行业启示

  • 本地部署选型策略:消费级硬件上应优先选择1B-2B参数的高效架构(如gemma3、llama3.2),而非盲目追求大参数模型,可在能效与性能间取得最佳平衡
  • 量化与架构优化价值:模型架构设计和量化策略对能效的影响超过参数量,建议结合具体应用场景选择最优配置,推动绿色AI落地
  • 基准测试标准完善:现有LLM基准多聚焦精度,需纳入能耗维度建立多目标评估体系,为本地部署提供全面的决策依据

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Inference 推理 Benchmark 基准测试 GPU GPU Deployment 部署