RTX Spark Slow on Linux? (2026): Unplug It or Use WSL
On native Linux, an RTX Spark laptop ran gpt-oss 120B at 13.9 tok/s, against 65.4 on Windows on the same machine. The cause is power management, not Linux itself: on AC power the GPU clock drops to about 780 MHz. Unplug the laptop and Linux runs at 61-67 tok/s. WSL runs at 65.1. Here is each symptom, its cause, and the measured fix.
Bottom Line
Native Linux on an RTX Spark laptop runs at about one fifth of the Windows speed when plugged in. The cause is a power-management component that only Windows can read. Two workarounds give back the speed today.
| Symptom on native Linux | Cause | Fix | Measured result |
|---|---|---|---|
| gpt-oss 120B at 13.9 tok/s (Windows: 65.4) | GPU clock held at ~780-790 MHz on AC power | Unplug the laptop | ~61-67 tok/s in LM Studio, clock ~1,800 MHz |
| Same slowdown | NPCF power table is encrypted; no Linux equivalent | Run the model in WSL on Windows | 65.1 tok/s on gpt-oss 120B |
| vLLM at 39.5 tok/s across 8 clients | Same clock cap | Run vLLM in WSL | 127 tok/s across 8 clients |
| Linux sees 28GB of a 128GB laptop | Firmware reserves 96GB for the GPU | Set the GPU reservation to 4GB from Windows | Linux sees 120GB |
| WSL runs out of memory on a big model | WSL gets 14GB by default | Raise memory= in .wslconfig | Load size is then set by your limit |
All figures come from one tester (MindStudio, MSI Prestige RTX Spark laptop, published 1 November 2026) unless marked otherwise. We did not test an RTX Spark machine.
What the headlines got wrong. The story spread as “inference slows to 35% on Linux.” The measured numbers say something else. 13.9 / 65.4 is 21%, not 35% (our arithmetic). And Linux is not slow in general: unplugged, it ran at about the Windows speed. This is a plugged-in power-state problem, so the fix is about power, not about the distro or the driver version.
What is happening
The RTX Spark laptop chip shares one memory pool between CPU and GPU, and one power budget between them. On Windows, a firmware component called the NVIDIA Platform Controllers and Framework (NPCF) sets how that budget is split. MindStudio reports three things about it:
- NPCF is Windows-only. It has no Linux equivalent.
- Its power table is encrypted. Only Windows can decrypt it.
- Disabling NPCF on Windows reproduces the slowdown. With NPCF off, one run fell from about 83 to about 31 tok/s. Turning it back on restored the speed.
Without NPCF, the Linux system falls back to a conservative power state when it is plugged in. MindStudio measured the clock at about 780-790 MHz under load on AC power. On Windows the same GPU ran near 2.2 GHz. A long benchmark took about 8 minutes on Windows and 37 minutes on Linux.
A developer on NVIDIA’s forum reported that on an affected unit “the EC publishes 0W power limits and the SoC falls back to 20W”, and asked for owner keys. NVIDIA replied: “We will not be providing keys for installing custom FW.” Both quotes come from Byteiota’s report; we did not read the forum thread.
Fix 1: unplug the laptop
On battery, the Linux system did not hold the low clock. MindStudio measured about 1,800 MHz under load and about 61-67 tok/s in LM Studio, against about 15 tok/s plugged in.
Use this to check that your setup works, or for a short session. It is not a fix for an always-on agent or a server, because the battery runs down.
Fix 2: run the model in WSL
WSL on the same laptop matched Windows: 65.1 tok/s on gpt-oss 120B against 65.4 natively, and vLLM at 127 tok/s across eight clients against 39.5 on native Linux.
WSL gets 14GB of memory by default. Raise it before you load a large model. In %USERPROFILE%\.wslconfig:
[wsl2]
memory=96GBPick the value from your own model size plus room for the KV cache. gpt-oss 120B needs about 61GB loaded at the 96GB GPU split (see Best Local LLM for RTX Spark). Then restart WSL with wsl --shutdown.
Fix 3: give Linux its memory back
On the MSI Prestige that MindStudio tested, the firmware reserved 96GB of the 128GB pool for the GPU, and native Linux saw only 28GB. The split can be changed only from Windows.
- Boot Windows and set the GPU memory reservation to 4GB in the vendor tool.
- Reboot into Linux. MindStudio reports Linux then saw 120GB, with a full 4K/120Hz desktop, audio, touchpad, keyboard and CUDA working.
MindStudio did not publish a large-model tok/s figure at the 4GB setting. The clock cap from Fix 1 still applies when you are plugged in. Other RTX Spark laptops ship with different default splits; StorageReview measured 62.5GB GPU-usable on a Surface Laptop Ultra at its default.
The distro and driver
MindStudio’s working install was Omarchy (Arch-based), whose kernel already bundles NVIDIA’s open-source driver. A Debian attempt came first and did not get there. Neither path fixes the clock cap. No driver version is reported to remove it.
Find your own number
Your figure will differ by laptop model, runtime, quant, and power mode. Measure it before you plan around it:
- Load the same model and prompt on Windows and on Linux.
- Run it plugged in, then on battery. Note tok/s each time.
- On Linux, watch the GPU clock during generation with
nvidia-smi -q -d CLOCK. - If the plugged-in clock sits near 800 MHz and the battery clock is far higher, you have this problem.
If you need native Linux, buy for it
If your plan is a Linux inference box, the laptop is the wrong form for it today. The DGX Spark uses the same GB10-class chip, ships with Linux (DGX OS), and ran gpt-oss 120B at 61 tok/s in the same MindStudio comparison. DGX Spark vs Strix Halo compares it with the cheaper 128GB route, which runs Linux with ROCm or Vulkan. For a machine list by memory tier, see the gear list.
If you already own the laptop, use WSL for long runs and wait for a driver fix. NVIDIA has published no timeline as of 6 November 2026.
Sources
- MindStudio: Linux on the NVIDIA RTX Spark laptop (1 November 2026: MSI Prestige, tok/s on Windows, WSL and Linux, clocks, NPCF, memory split; single tester)
- Hacker News: Nvidia RTX Spark encrypted firmware slows inference speed to 35% on Linux (4 November 2026; 614 points and 289 comments when read 6 November)
- Byteiota: Linux cuts AI inference to 35% (4 November 2026; source of the NVIDIA forum quotes. Its headline says DGX Spark, but the article is about the RTX Spark laptop)
- daily.dev: RTX Spark laptops are fast, until Linux and Arm get involved (4 November 2026)
- StorageReview: Surface Laptop Ultra for Business review (15 October 2026: 62.5GB GPU-usable at the default split)
See Also
- Best Local LLM for RTX Spark (2026) — which models to run, and why the GPU split must go to 96GB
- Best Models for NVIDIA DGX Spark — the desktop GB10 box that runs Linux natively
- DGX Spark vs Strix Halo — the two 128GB Linux routes
- Is the DGX Spark Worth It? — the price case for the desktop version
- llama.cpp Flags Explained — the flags to set once the clock problem is solved