Nvidia's RTX Spark laptops — shipping from MSI, Asus, and Dell — pair an ARM CPU with a CUDA-capable GPU and up to 128GB of unified memory, positioning them as portable AI development machines. Linux installs, boots, and runs a full desktop with GPU acceleration through Nvidia's open kernel driver. That much works. The problem is what happens when you plug in the power cable. On wall power, the GPU clock speed collapses from roughly 2.2 GHz to 780 MHz under Linux. The result: inference on GPT-OSS-120B drops from 65.4 tokens per second on Windows to 13.9 on native Linux — a 4.7× performance gap. On battery, Linux clocks climb back to ~1,800 MHz and performance approaches Windows levels. The throttling is not thermal; GPU temperatures stay between 40°C and 47°C. The bottleneck is firmware. The culprit is a component called the Nvidia Platform Controllers and Framework, visible in Windows Device Manager but entirely absent on Linux. This firmware layer manages shared power delivery between CPU and GPU — a necessity in a laptop form factor where both chips draw from the same power budget. Disabling it in Windows reproduces the exact same clock collapse, confirming it as the root cause. The firmware's power table is encrypted, and only the Windows driver holds the decryption key. Installation itself requires multiple workarounds. Debian's vanilla kernel lacks the necessary drivers; Omarchy (an Arch-based distro) ships Nvidia's open driver and works after manual configuration. Wi-Fi fails during install, requiring a USB Ethernet adapter. The laptop's firmware reserves 96GB of unified memory for the GPU by default, and that allocation is only adjustable from Windows. Without changing it, Linux sees only 28GB of the machine's 128GB. HDMI output doesn't work during early boot, forcing installers to point cameras at the laptop screen. WSL emerges as the practical workaround. Running atop Windows' driver stack, it inherits full GPU boost behavior and benchmarks at 65.1 tokens per second — essentially matching native Windows. With a manual memory allocation bump from WSL's default 14GB, Linux-first tools like vLLM and SGLang run at near-full speed. For developers who need Linux tooling today, WSL on these machines delivers what native Linux cannot. The deeper issue is architectural. Nvidia built a unified-memory ARM platform capable of running 120B-parameter models locally, then shipped every laptop variant with a firmware dependency that only Windows can satisfy. The DGX Spark desktop runs Nvidia's own Linux build natively, proving the hardware supports it. The laptop's power management layer is the sole barrier, and it's a barrier Nvidia created and only Nvidia can remove. A real fix requires Nvidia to port the platform controller's power management logic into its Linux driver — effectively making the laptop a portable DGX Spark. Until then, developers face a choice between native Linux at one-fifth speed or WSL as a capable but philosophically unsatisfying middle ground. The hardware is open. The firmware is not.
