Developers hoping to escape Windows on Nvidia's RTX Spark laptops just hit a wall. A detailed hands-on review published by MindStudio found that while Linux does install and run on these AI-focused machines — with GPU acceleration, a full desktop, and working peripherals — encrypted power-management firmware locks the GPU to roughly a third of its Windows clock speed, collapsing local AI inference from 65.4 tokens per second to just 13.9 on the same hardware and model.
That is a gap of nearly five times on the exact machines marketed at AI developers. For more context on this story, see our ongoing latest AI developments.
The benchmarks: one model, four configurations
The review benchmarked GPT-OSS-120B, a 120-billion-parameter open model that fits on the laptop's 128GB of unified memory, across four configurations:
- Windows on ARM (native): 65.4 tokens per second
- WSL (Linux inside Windows): 65.1 tokens per second
- Nvidia DGX Spark (desktop unit): 61 tokens per second
- Native Linux (same laptop): 13.9 tokens per second
The pattern is striking. WSL — Linux running as a subsystem on top of Windows — matches native Windows almost exactly. The desktop-class DGX Spark is slightly slower than the laptop on Windows, likely reflecting its different thermal profile. But native Linux on the same laptop manages barely a fifth of its Windows performance.
Encrypted firmware is the root cause
The culprit is not the open-source Nvidia kernel driver, which works and provides CUDA acceleration. It is a Windows-only component called the Nvidia Platform Controllers and Framework, which manages power delivery between the CPU and GPU. The review found this firmware has no Linux equivalent — and its power table is encrypted, so only the Windows driver can read it.
The practical effect: once the laptop is plugged into wall power, the GPU gets locked to roughly 780 MHz on Linux instead of the 2.2 GHz-plus it reaches on Windows. On battery, the difference narrows, but plugged-in operation is exactly when you would expect peak inference performance — and exactly when Linux cannot deliver it.
The memory-split gotcha
Clock speed is not the only Windows-dependent setting. The laptop's firmware reserves 96GB of its 128GB unified memory for the GPU by default, and — again — only Windows exposes a setting to change that split.
The reviewer discovered this the hard way: after getting Linux to boot, only 28GB of the 128GB was visible to the system. The fix required booting back into Windows, dropping the GPU memory reservation to 4GB, and then rebooting into Linux, which could then see 120GB. Without that detour, large models that should fit comfortably on the machine are simply out of reach on Linux.
Installing Linux is still an adventure
Even getting to the throttled GPU was non-trivial. The review's first attempt with Omarchy, an Arch-based distribution that already supports the chip on Asus and Dell hardware, failed on the MSI Prestige N16 Flip AI Plus with a flickering, unstable display during boot. Switching to Debian worked more reliably but demanded workarounds: the installer could not connect over Wi-Fi, requiring a USB Ethernet adapter, and a vanilla kernel shipped without support for the laptop's keyboard, touchpad, or GPU.
Returning to Omarchy solved the driver problem, since its kernel bundles Nvidia's open-source driver. Early in the process, HDMI output also refused to work during installation — the reviewer resorted to pointing an external camera at the laptop's internal screen just to follow the installer. Once configured, the final system ran a full desktop at 4K and 120Hz with working audio, keyboard, touchpad, and CUDA acceleration.
Why this matters for local AI developers
The RTX Spark platform ships in machines including the MSI Prestige, the Asus ProArt, and the Dell XPS, all using the same GPU design as Nvidia's DGX Spark desktop unit. But there is a key divergence: the DGX Spark ships with Nvidia's own Linux build out of the box, while every laptop variant ships with Windows on ARM. Much of the serious local-AI tooling — vLLM, SGLang — is built Linux-first, which is precisely why developers buying these machines want Linux on them.
The review concludes that WSL is the practical answer today: it rides the Windows driver stack, keeps full GPU boost whether plugged in or on battery, and still runs Linux-first AI tooling. Native Linux remains a hobbyist project — functional, but leaving most of the hardware's advertised performance on the table.
A real fix depends on Nvidia porting the platform controller's power-management logic into its Linux driver. Until that happens, the review notes, the laptop cannot become what buyers want it to be: a portable DGX Spark. Nvidia has not publicly commented on the finding or signaled a timeline for Linux support. For developers deciding between these laptops and the desktop unit, the current state of the firmware is a material caveat — the spec sheet says 2 GHz, but on Linux, only Windows knows how to ask for it.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
