Benchmarking 15 "E-Waste" GPUs with Modern Workloads
141 points
• 4 days ago
• Article
Link
退役的企业级 GPUs 常被视为电子垃圾，但对 homelab 爱好者来说，它们是成本极低的显存来源。像 K80 、 P100 、 V100 这类卡能以原价的一小部分入手，因此在 AI 推理和 3D 渲染等现代任务中仍很有吸引力。尽管这些设备官方已退役、缺乏最新驱动支持，但通过使用较旧的软件版本或容器化环境，可以在不频繁更新的情况下，继续满足个人项目的需要。
基准测试采用容器化方法，覆盖了从 ResNet50 的图像训练、 Vision Transformer 的推理，到 LLM 的文本生成与科学计算等多种负载。跨几代显卡的测试显示，性能大体随发布日期提升，且 V100 表现尤其突出、性价比高。另一方面，虽然功耗效率对于持续高可用场景是个问题，但 homelab 的间歇性工作模式使得能耗折衷对许多用户而言是可接受的。
多 GPU 扩展的实验结果令人鼓舞：大多数任务随着显卡数量增加呈线性性能提升。虽会有一定的通信开销，但测试表明在标准的 4U 机箱内塞满这些卡是可行的，在典型的 homelab 配置下并未出现严重的收益递减。单节点内混插不同代显卡也是可行的，但在某些应用中弱卡可能成为瓶颈。
CPU 的选择也有微妙影响：整体上更快的单核性能能带来更好的表现；但对于像 Whisper 或某些 Transformer 工作负载，更多的核心数有时反而会导致轻微的性能下降。不同工作站级主板上的测试结果大致一致，表明面向预算的基于 X99 的硬件完全能够满足希望构建高密度高性能 GPU 节点的需求。
综合来看，V100 是需要同时处理 Whisper 语音识别、 LLM 和图像分析等混合工作负载时的首选。将这类显卡与一颗 8 核 CPU 及可靠的工作站主板结合，用户可以搭建出应对复杂任务的强大服务器，而无需承担现代企业级硬件的高昂成本。这些发现验证了对退役企业设备的再利用价值，说明"过去的"硬件依然能满足当下的计算需求。
Decommissioned enterprise GPUs, often labeled as e-waste, represent a highly cost-effective source of VRAM for homelab enthusiasts. Cards such as the K80, P100, and V100 can be acquired for fractions of their original cost, making them attractive for modern tasks like AI inference and 3D rendering. While these units are officially end-of-life and lack contemporary driver support, utilizing older software builds or containerized environments allows them to remain fully functional for personal projects without the need for constant updates.
The benchmarking process utilized a containerized approach, running diverse workloads ranging from ResNet50 image training and Vision Transformer inference to LLM text generation and scientific computing. By testing cards across several generations, it became clear that performance generally scales with release date, though the V100 stands out as a particularly high-value performer. Additionally, while power efficiency is a concern for constant high-availability use, the intermittent nature of homelab workloads makes the energy trade-off manageable for many users.
Experiments with multi-GPU scaling revealed encouraging results, as most tasks showed linear performance gains as more cards were added to the system. While some communication overhead is expected, the tests suggest that filling a standard 4U chassis with these GPUs is a viable strategy, as there is little evidence of severe diminishing returns in typical homelab configurations. Mixing GPU generations within a single node is also possible, though performance may be bottlenecked by the weaker cards in specific applications.
CPU selection plays a nuanced role in these setups, with faster single-core performance generally providing better results across the board. However, for specialized tasks like Whisper or certain Transformer workloads, higher core counts can occasionally lead to slight performance degradation. Testing across different workstation-grade motherboards showed consistent results, suggesting that budget-friendly X99-based hardware is perfectly adequate for those looking to assemble a dense, high-performance GPU node.
Ultimately, the V100 emerges as a top choice for projects requiring a blend of Whisper speech recognition, LLM capabilities, and image analysis. By combining such cards with an 8-core CPU and a reliable workstation motherboard, users can build powerful servers that satisfy complex workloads without the expense of modern enterprise hardware. These findings validate the potential of repurposed enterprise gear, proving that yesterday's hardware remains a potent resource for today's computing demands.
64 comments • Comments Link
• Tesla P4 是一个颇具吸引力的预算级推理选项，提供 8GB VRAM 和 75W TDP，价格约为 $80，但其提示词吞吐率明显低于现代消费级显卡。
• 在紧凑机箱中装入多张企业级显卡需要定制散热方案，并带来巨大的功耗挑战；根据当地电价，某些配置的月度电费可能高达数百美元。
• Radeon Pro V620 以有竞争力的价格提供 32GB VRAM，且仍受当前 ROCm 版本支持，是替代老旧 Nvidia 服务器硬件的更现代选择。
• Intel 的 B70 目前受限于尚不成熟且功能有限的软件栈，通常需要专有分支，且在性能吞吐上难以与 AMD 的 ROCm 或 Nvidia 的 CUDA 相抗衡。
• BC-250（源自 PS5 的定制芯片）在 Ethereum 挖矿衰退后成为一种新颖的低成本推理选择，且在 GitHub 上有活跃的社区开发。
• 企业级 GPU 在高负载下通常以接近最大 TDP 的状态运行，因此基准测试应依赖精确的功耗监测，而不是仅看标称峰值规格。
• 将系统扩展到 16 张及以上 GPU 会带来极端的架构要求，包括高压供电和潜在的 PCIe 通道瓶颈，使关注点从单纯追求性能转向追求具有成本效益的硬件密度。
• 老旧企业级显卡（如 K80）常因多年热循环而出现可靠性问题，这引发了对当前高端 AI 硬件长期可行性的担忧。
• 旧款 GPU 在软件支持方面面临重大障碍：它们依赖的专有二进制 blob 可能与现代 Linux 内核不兼容，这一点有别于那些擅长光线追踪或视频编码的现代媒体导向显卡。
• 基准测试应优先考虑实际可用性指标，如 tokens-per-second 、针对 Qwen 27B 等流行模型的 VRAM 容量以及上下文窗口限制，以帮助预算有限的用户做出更明智的硬件选择。
本次讨论集中在自托管 AI 推理中旧企业级硬件与现代消费级 GPU 的权衡。虽然像 Tesla P4 或 K80 这样的旧卡以极低价格提供较高的 VRAM 密度，但参与者指出它们往往伴随显著的隐性成本，包括更高的功耗、额外的散热需求以及逐渐减少的软件支持。共识倾向于在硬件采购成本与最终系统"可用性"之间寻找平衡，许多用户更愿意针对需要大量内存的特定大型语言模型进行优化。总体来看，社区更偏好重新利用专用或退役硬件，这反映出一种由发烧友驱动、旨在规避当代 AI 设备高准入门槛的工程趋势。 • The Tesla P4 is a compelling budget option for inference, offering 8GB of VRAM and 75W TDP for roughly $80, though prompt ingestion speeds are notably slower than modern consumer cards.
• Fitting multiple enterprise cards into a compact case requires custom cooling solutions and presents significant power consumption challenges, with some setups potentially costing hundreds of dollars per month in electricity depending on local utility rates.
• Radeon Pro V620 cards offer 32GB of VRAM at a competitive price point and remain supported by current ROCm releases, making them a more modern alternative to aging Nvidia server hardware.
• Intel's B70 GPU is currently hampered by a maturing but limited software stack, often requiring proprietary forks and failing to achieve full performance throughput compared to AMD's ROCm or Nvidia's CUDA.
• The BC-250 (a custom chip derived from the PS5) has emerged as a novel, low-cost option for inference following the decline of Ethereum mining, with active community efforts documented on GitHub.
• Enterprise GPUs frequently operate near their maximum TDP during intensive workloads, necessitating accurate power monitoring for effective benchmarking rather than relying solely on peak specifications.
• Scaling to 16+ GPUs introduces extreme infrastructure requirements, including high-voltage power delivery and potential PCIe lane bottlenecks, shifting the goal from pure performance to cost-effective hardware density.
• Aging enterprise cards, such as the K80, often suffer from reliability issues caused by years of thermal cycling, raising questions about the long-term viability of current high-end AI hardware assets.
• Older GPUs face significant hurdles regarding software support, as they rely on proprietary binary blobs that may not be compatible with modern Linux kernels, unlike modern media-focused cards that excel in ray tracing or video encoding.
• Benchmarking efforts should prioritize real-world usability metrics like tokens-per-second, VRAM capacity for popular models like Qwen 27B, and context window limitations to help budget users make informed hardware choices.
The discussion centers on the trade-offs between legacy enterprise hardware and modern consumer-grade GPUs for self-hosted AI inference. While older cards like the Tesla P4 or K80 offer high VRAM density for a very low price, participants noted that they often come with significant hidden costs, including power consumption, cooling requirements, and dwindling software support. The consensus favors finding a balance between hardware cost and the "usability" of the resulting system, with many users looking to optimize for specific large language models that require significant memory overhead. Ultimately, the community shows a preference for repurposing specialized or retired hardware, reflecting a broader trend of enthusiast-driven engineering aimed at bypassing the high entry barrier of current-generation AI equipment.