搭载 Nvidia RTX Spark 芯片的笔记本电脑在运行 Linux 系统并连接电源时,遭遇了严重的 AI 模型推理性能下降,其速度仅为 Windows 或 WSL 环境下的约 35% 1。这一瓶颈导致 GPU 时钟频率被锁定在约 780 MHz,而同等条件下 Windows 系统中的频率可达约 2.2 GHz 1。根本原因在于 Linux 系统缺乏 Windows 独有的“Nvidia Platform Controllers and Framework”电源管理固件组件,该组件管理的 CPU/GPU 共享电源预算表在 Linux 中因无法解密读取而失效 1。
具体的基准测试数据显示,在运行 GPT-OSS-120B 模型时,Windows 环境的吞吐量达到 65.4 tokens/s,WSL 环境为 65.1 tokens/s,而原生 Linux 环境下仅为 13.9 tokens/s 1。值得注意的是,当断开电源后,Linux 系统的性能可恢复正常水平(约 61-67 tokens/s),说明降频现象仅出现在插电状态 1。此外,在统一内存配置下,默认有 96GB 内存被保留给 GPU,用户需在 Windows 端进行调整才能使 Linux 访问更多 RAM 1。目前,彻底解决该问题依赖于 Nvidia 将平台控制器的电源管理逻辑移植到 Linux 驱动中,在此之前,WSL 被视为更实用的替代方案 1。
Laptop users running Linux on devices equipped with Nvidia’s RTX Spark chips are experiencing drastic performance drops in AI model inference when their systems are connected to a power source. The issue stems from the GPU clock speed being throttled to approximately 780 MHz while plugged in, a significant reduction compared to the roughly 2.2 GHz speeds achieved under Windows 1. Benchmark tests using the GPT-OSS-120B model illustrate the severity of this bottleneck: inference rates plummeted to just 13.9 tokens per second on native Linux, whereas Windows delivered 65.4 tokens per second and Windows Subsystem for Linux (WSL) managed 65.1 tokens per second 1. Notably, once the laptop is disconnected from its power adapter, Linux performance returns to normal levels, ranging between 61 and 67 tokens per second 1.
The root cause of this degradation lies in the absence of critical power management firmware components on Linux that are exclusive to Windows. Specifically, the "Nvidia Platform Controllers and Framework" manages a shared power budget table for the CPU and GPU, but this encrypted logic cannot be decrypted or read by Linux drivers 1. Consequently, the operating system fails to properly regulate power states during AC connection, leading to the locked clock speeds. While WSL currently offers a more practical workaround for these users, a definitive solution requires Nvidia to update its Linux drivers to support the platform controller's encrypted power management logic 1. Additionally, in unified memory configurations, 96GB is reserved for the GPU by default, meaning users must adjust settings within Windows to allow Linux access to more RAM 1.
评论
还没有评论,欢迎留下第一条。