Linux makes 14 tokens a second on a 120B model on Nvidia RTX Spark
On wall power, Linux is about five times slower than Windows. On battery, Linux has about the speed of Windows.
A test of the MSI Prestige N16 Flip AI Plus, a laptop with the Nvidia RTX Spark chip, found a large Linux slowdown. On wall power, the graphics chip stays near 780 MHz on Linux, against more than 2 GHz on Windows. On a 120 billion parameter model, Windows made 65.4 tokens a second and Linux made 13.9. A token is a small piece of text.
How it works
A Windows only part, the Nvidia Platform Controllers and Framework, shares power between the CPU, the main chip, and the GPU, the graphics chip. The firmware gives it an encrypted table of power adjustments, and only the Windows driver can decrypt it. When the test team turned this part off in Windows, the same slowdown occurred.
The numbers
On the 120B model, Windows made 65.4 tokens a second and WSL made 65.1. WSL is Windows Subsystem for Linux, a Linux environment in Windows. Native Linux made 13.9. A long test took 8 minutes on Windows and 37 minutes on Linux. Heat was not the cause.
What is not known
No plan from Nvidia for a repair is known. A full repair needs Nvidia to move the power logic into its Linux driver. Setup is also hard. Only Windows can change the 96GB memory division for the GPU.
Insight: The fastest Linux on this laptop is in Windows
SUSE, a Linux company, makes an agent platform on Nvidia chips, and ASRock Rack makes Nvidia racks. But on this laptop, the fastest Linux is in Windows, as WSL. Nvidia holds the key to the power table, and partners must have it for speed. This looks like a toll gate that Nvidia can open or keep shut. Test your own tools on the same machine.
Sources
Posted
