RTX Spark laptops run Linux inference at a fifth of Windows speed when plugged in: Reveneau

Why it mattersPicking the OS for your inference box is now a firmware question. WSL on Windows runs faster than native Linux on these laptops until Nvidia adds the power logic for Linux.

A Nvidia RTX Spark laptop running native Linux hits a fifth of Windows speed. That slowdown turns up only when the laptop is plugged into wall power. That is the finding of a post on the MindStudio blog. The post reached 146 points on Hacker News.

The test machine is an MSI Prestige N16 Flip AI Plus. The platform is sold in machines from MSI, Asus and Dell. All of them use the same GPU as Nvidia's desktop DGX Spark unit, which does not have the same problem. The laptops come with Windows on ARM and CUDA support.

What the benchmark shows

MindStudio ran GPT-OSS-120B on the GPU. Windows hit 65.4 tokens per second. WSL, the Linux environment inside Windows, hit 65.1. The DGX Spark desktop, running Nvidia's own Linux build, hit 61. Native Linux on the laptop, plugged in, hit 14.

The post also ran vLLM with eight concurrent clients. WSL reached 127 tokens per second. Native Linux reached 39.5. A longer run told the same story in time: Windows finished in about 8 minutes near 2.2 GHz. Native Linux took 37 minutes at 780 MHz.

Oddly, the direction of the problem flips with the power source. Unplugged, native Linux hit 61 to 67 tokens per second on LM Studio, close to Windows. Plugged in, the same test dropped to about 15. The GPU clock on native Linux was 1,800 MHz on battery and 786 MHz on wall power. Windows held the clock near 2 GHz in both cases.

The cause, named

MindStudio traces the gap to a Windows-only component called the Nvidia Platform Controllers and Framework. It manages the power budget between CPU and GPU on the laptop. The desktop DGX Spark does not have to do that. Disabling it in Windows reproduced the slowdown: tokens per second dropped from 83 to 31, and the clock fell from 2 GHz to 786 MHz. Turning it back on restored normal speed.

The deeper problem is that the firmware hands this controller a table of power settings that is encrypted. Windows has the key. Linux does not. MindStudio tried a kernel patch and triggered a kernel panic.

The practical answer, this month

For now, WSL is the answer. It runs on top of the Windows driver stack. So it keeps the Windows boost logic, plugged in or on battery. Linux-first tools like vLLM and SGLang run inside it. One manual change is needed. WSL's default memory cap is 14 GB. The laptop has 128 GB of unified memory, so the cap must be raised.

Dual-booting a native Linux is possible. MindStudio used the Omarchy distribution, which bundles Nvidia's open kernel driver for this chip. Debian worked too, after manual driver and network workarounds, including a USB Ethernet adapter during install. Native Linux still throttles on wall power. The full fix has to come from Nvidia putting the platform controller logic into its Linux driver.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.