Hardware / ARTICLE

Nvidia RTX Spark Laptop Locks GPU at 780 MHz on Linux When Plugged In

On RTX Spark laptops such as the MSI Prestige, Linux installs with GPU acceleration, but when plugged in the GPU drops to around 780 MHz and inference throughput falls to roughly one-fifth of Windows.

Installation Requires Manual Steps

Linux runs on RTX Spark laptops like the MSI Prestige N16 Flip AI Plus, with GPU acceleration through Nvidia’s open kernel driver. The setup is not straightforward. An initial attempt using Omarchy produced flickering and unstable display on the MSI Prestige. Debian installed more reliably, but the installer had no Wi-Fi, requiring a USB Ethernet adapter, and the vanilla kernel did not recognize keyboard, touchpad, or GPU. Returning to Omarchy fixed driver support because its kernel already included Nvidia’s open-source driver. At that point Linux saw only 28 GB of the 128 GB unified memory; the firmware reserved 96 GB for the GPU, and that split could only be changed from Windows. Reducing the Windows GPU reservation to 4 GB let Linux see about 120 GB. After those changes, the laptop drove a 4K display at 120 Hz with working audio, keyboard, touchpad, and CUDA acceleration. HDMI output was unavailable during early installation, so remote steps required pointing an external camera at the internal screen.

Plugged-in Linux Falls to a Fraction of Windows Speed

Benchmarks with GPT-OSS-120B show the gap. Windows reached 65.4 tokens per second, WSL 65.1, and the DGX Spark desktop 61. Native Linux on the same laptop managed 13.9. In a vLLM test with eight concurrent clients, WSL hit 127 tokens per second while native Linux reached 39.5. A longer run finished in about 8 minutes on Windows with the GPU near 2.2 GHz; Linux took 37 minutes with the GPU held around 790 MHz. Temperatures stayed between 40 and 47°C, so thermal throttling does not explain the drop. The slowdown also reverses with power state. Unplugged, Linux ran LM Studio at 61–67 tokens per second, close to Windows. Plugged in, the same test fell to about 15 tokens per second. Clock speeds tracked the change: about 1800 MHz unplugged under load, dropping to roughly 780 MHz on wall power.

Windows-Only Firmware Controller Causes the Limit

The Nvidia Platform Controllers and Framework appears in Windows Device Manager. It manages a shared power budget between CPU and GPU, a job not needed on the desktop DGX Spark where the GPU has its own power path. Disabling that controller in Windows reproduced the Linux behavior: token throughput dropped from about 83 to about 31, and GPU clock fell from about 2 GHz to 786 MHz. Re-enabling it restored full speed immediately. The controller is the limit, not the kernel, distribution, or GPU driver. The firmware supplies the controller with an encrypted power settings table. Windows holds the decryption key; Linux does not. Software attempts to enable the controller on Linux failed, and one triggered a kernel panic. Extracting the key from the Windows driver would be a fragile single-machine hack. A fix requires Nvidia to port the platform controller’s power management logic into its Linux driver.

WSL Works as a Practical Substitute

WSL sits on the Windows driver stack and inherits full GPU boost behavior on battery or wall power. It runs Linux-first tools such as vLLM and SGLang without the clock throttling that affects native Linux. The default WSL memory allocation of 14 GB is too small for large models on a 128 GB machine and must be increased manually. After that adjustment, WSL benchmarks land close to native Windows performance, making it the more practical choice until Nvidia ships a Linux-side fix.

END
Feedback