Research Papers 论文研究 4d ago Updated 3d ago 更新于 3天前 49

DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs DumpsterCluster:从回收旧GPU到用60美元显卡服务LLaMA-70B

A 128-GPU "DumpsterCluster" was physically built from second-hand components at a total cost of $22K, compared to $600K for an 8-GPU B200 system, demonstrating massive cost savings for LLM inference infrastructure. Pipeline-parallel optimizations enabled a V100-based DumpsterCluster to achieve competitive LLaMA-70B throughput, validating the production viability of repurposed hardware for modern large language model serving. Older GPUs consume significantly more energy per token, making total co 用二手V100 GPU构建128卡集群仅需2.2万美元,相比8卡B200系统(60万美元)成本优势显著 通过流水线并行优化,V100集群可实现有竞争力的LLaMA-70B推理吞吐量 二手GPU每token能耗更高,仅在电价低廉地区具备经济可行性 二手系统碳排放比新一代硬件高4倍(8B模型)至40倍以上(70B模型) GPU再利用必须与低碳能源结合才能实现环境可持续性

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • A 128-GPU "DumpsterCluster" was physically built from second-hand components at a total cost of $22K, compared to $600K for an 8-GPU B200 system, demonstrating massive cost savings for LLM inference infrastructure.
  • Pipeline-parallel optimizations enabled a V100-based DumpsterCluster to achieve competitive LLaMA-70B throughput, validating the production viability of repurposed hardware for modern large language model serving.
  • Older GPUs consume significantly more energy per token, making total cost of ownership favorable only in regions with inexpensive electricity, creating a critical economic dependency on local energy pricing.
  • Carbon emissions from second-hand systems are approximately 4x higher for 8B models and over 40x higher for 70B models compared to current-generation hardware under grid-average carbon intensity, revealing severe environmental trade-offs.
  • GPU repurposing is not universally sustainable; it must be strategically coupled with low-carbon energy sources to be both economically viable and environmentally responsible.

Why It Matters

This research directly addresses the growing sustainability and affordability crisis in AI infrastructure as datacenters retire functional GPUs. For AI practitioners and organizations seeking to reduce inference costs, it provides empirical evidence that second-hand hardware can serve modern LLMs at a fraction of the price—but only under specific energy and geographic conditions. The findings challenge the assumption that hardware reuse is inherently green and offer a data-driven framework for evaluating when repurposing makes sense.

Technical Details

  • The DumpsterCluster consists of 128 second-hand GPUs (V100-based) assembled entirely from used components, running continuously for one year to validate long-term production viability.
  • Pipeline-parallel optimization techniques were employed to distribute LLaMA-70B inference workloads across the heterogeneous, older GPU cluster, achieving competitive token throughput despite the hardware's age.
  • Economic analysis compares the $22K DumpsterCluster against a $600K 8-GPU B200 system, highlighting a roughly 27x cost advantage for the repurposed cluster.
  • Environmental impact was measured in terms of carbon emissions per token, revealing that under average grid carbon intensity, second-hand hardware produces ~4x more emissions for 8B models and >40x more for 70B models relative to modern GPUs.
  • The study introduces a context-dependent viability framework where total cost of ownership and sustainability are jointly determined by local electricity pricing and grid carbon intensity.

Industry Insight

  • Organizations in regions with cheap, clean electricity (e.g., hydro-rich or nuclear-powered grids) should prioritize second-hand GPU procurement as a cost-effective strategy to expand inference capacity without proportional carbon cost increases.
  • AI providers serving large models (70B+) should be cautious about repurposing older hardware in carbon-intensive grids, as the environmental penalty scales dramatically with model size—potentially undermining ESG commitments and regulatory compliance.
  • The paper suggests a emerging market opportunity: datacenter operators and cloud providers could build or broker "green repurposing" clusters specifically sited in low-carbon, low-cost energy regions, creating a new niche in sustainable AI infrastructure.

TL;DR

  • 用二手V100 GPU构建128卡集群仅需2.2万美元,相比8卡B200系统(60万美元)成本优势显著
  • 通过流水线并行优化,V100集群可实现有竞争力的LLaMA-70B推理吞吐量
  • 二手GPU每token能耗更高,仅在电价低廉地区具备经济可行性
  • 二手系统碳排放比新一代硬件高4倍(8B模型)至40倍以上(70B模型)
  • GPU再利用必须与低碳能源结合才能实现环境可持续性

为什么值得看

这篇论文首次通过实际部署验证了退役GPU集群服务现代LLM推理的可行性,为AI基础设施的成本优化和可持续发展提供了重要参考。研究揭示了硬件再利用的经济与环境权衡,对AI从业者的采购决策和绿色计算策略具有指导意义。

技术解析

  • 构建128卡V100 DumpsterCluster,总成本2.2万美元,全部采用二手组件,实际运行一年验证稳定性
  • 采用流水线并行优化技术,在V100集群上实现LLaMA-70B模型的推理部署,达到有竞争力的吞吐量
  • 对比基准:8卡B200系统成本60万美元,凸显二手硬件的显著成本优势
  • 能耗分析显示旧GPU每token能耗显著高于新一代硬件,直接影响总拥有成本(TCO)

行业启示

  • 二手GPU集群可作为低成本AI推理的可行方案,但需优先考虑电价低廉的地区部署以抵消高能耗成本
  • 碳排放是硬件再利用的关键制约因素,绿色能源供应是必要条件,否则环境效益可能为负
  • AI基础设施决策需综合考量经济成本与环境影响,不能仅看硬件采购成本,应建立全生命周期评估框架

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。

GPU GPU LLaMA LLaMA Inference 推理 Deployment 部署 Research 科学研究