Updated on November 4, 2026
AI engineers purchasing Nvidia’s new 128GB unified-memory RTX Spark laptops—such as the MSI Prestige N16 Flip AI Plus, Asus ProArt, and Dell XPS—with plans to wipe Windows and run native Linux for local LLM inference are hitting a severe firmware wall. According to a hands-on hardware and driver investigation published by MindStudio Director of Product Luis Chavez-Mattos, booting native Linux on an RTX Spark laptop causes the GPU clock speed to collapse from 2.2 GHz to roughly 780 MHz the moment the machine is plugged into wall power, slashing local inference throughput by 78.7%.
The RTX Spark mobile platform pairs an ARM CPU with the same unified-memory GPU architecture found in Nvidia’s desktop DGX Spark workstation, offering up to 128GB of shared system and video memory capable of holding 120-billion-parameter models locally. However, while the desktop DGX Spark ships with Nvidia’s custom Linux OS out of the box, every RTX Spark laptop ships with Windows on ARM .
When MindStudio benchmarked GPT-OSS-120B across operating systems on the MSI Prestige RTX Spark laptop, Windows on ARM delivered 65.4 tokens per second and Windows Subsystem for Linux (WSL2) achieved 65.1 tokens per second—actually edging out the desktop DGX Spark (61.0 tokens per second). On native dual-boot Linux using Nvidia’s open kernel driver, however, the exact same laptop struggled to reach 13.9 tokens per second.
Tracing the clock-speed collapse revealed that dynamic power allocation between the ARM CPU and GPU is governed by a proprietary Windows-only component called the Nvidia Platform Controllers and Framework . Because the firmware’s power delivery table is cryptographically encrypted, only the Windows driver can decrypt it and authorize boost clocks above 780 MHz; native Linux cannot read the power table and defaults to a fallback safe clock.
Hardware Tested : Nvidia RTX Spark 128GB unified-memory laptop (MSI Prestige N16 Flip AI Plus, sharing silicon lineage with Asus ProArt, Dell XPS, and desktop DGX Spark), audited in the MindStudio RTX Spark Linux benchmark report .GPT-OSS-120B Throughput Gap : Native Windows hit 65.4 tok/s and WSL2 hit 65.1 tok/s, whereas native dual-boot Linux dropped 78.7% to 13.9 tok/s (compared to 61.0 tok/s on desktop DGX Spark).Root Cause (Encrypted Power Table) : GPU clock speed locks at ~780 MHz on AC wall power under Linux instead of ~2.2 GHz on Windows because the Nvidia Platform Controllers and Framework power table is encrypted and lacks a Linux driver port.96GB Firmware Memory Trap : By default, system firmware reserves 96GB of the 128GB unified RAM exclusively for the Windows GPU driver, leaving native Linux with only 28GB of visible system RAM unless reconfigured from Windows first.
Benchmark & Firmware Comparison Across RTX Spark Environments
For developers evaluating whether to run native Windows, WSL2, native Linux, or a desktop DGX Spark for local 70B–120B model serving, the MindStudio telemetry highlights how firmware locks dictate real-world performance:
Execution Environment
GPU Clock on Wall Power
GPT-OSS-120B Speed (tok/s)
Visible RAM Out-of-Box
Linux AI Tooling (vLLM / SGLang) Support
Windows on ARM (Native) ~2,200 MHz (Full Boost)
65.4 tok/s 128GB (96GB GPU / 32GB OS)
Limited (Windows ports required)
WSL2 on Windows on ARM ~2,200 MHz (Full Boost)
65.1 tok/s (99.5% of Native) Configurable up to 120GB+
Full (Rides Windows power driver)
Nvidia DGX Spark (Desktop) ~2,000+ MHz (Desktop Linux)
61.0 tok/s 128GB Unified
Full Native Linux Out-of-Box
Native Dual-Boot Linux (Laptop) ~780 MHz (Firmware Locked) 13.9 tok/s (-78.7%) 28GB (Until Windows split changed)Full, but severely clock-throttled
Windows on ARM (Native)
GPU Clock on Wall Power ~2,200 MHz (Full Boost)
GPT-OSS-120B Speed (tok/s) 65.4 tok/s
Visible RAM Out-of-Box 128GB (96GB GPU / 32GB OS)
Linux AI Tooling (vLLM / SGLang) Support Limited (Windows ports required)
WSL2 on Windows on ARM
GPU Clock on Wall Power ~2,200 MHz (Full Boost)
GPT-OSS-120B Speed (tok/s) 65.1 tok/s (99.5% of Native)
Visible RAM Out-of-Box Configurable up to 120GB+
Linux AI Tooling (vLLM / SGLang) Support Full (Rides Windows power driver)
Nvidia DGX Spark (Desktop)
GPU Clock on Wall Power ~2,000+ MHz (Desktop Linux)
GPT-OSS-120B Speed (tok/s) 61.0 tok/s
Visible RAM Out-of-Box 128GB Unified
Linux AI Tooling (vLLM / SGLang) Support Full Native Linux Out-of-Box
Native Dual-Boot Linux (Laptop)
GPU Clock on Wall Power ~780 MHz (Firmware Locked)
GPT-OSS-120B Speed (tok/s) 13.9 tok/s (-78.7%)
Visible RAM Out-of-Box 28GB (Until Windows split changed)
Linux AI Tooling (vLLM / SGLang) Support Full, but severely clock-throttled
The Hidden 96GB Unified Memory Split Trap on Dual-Boot Systems
Even before hitting the 780 MHz clock ceiling, engineers attempting to dual-boot Debian or Arch-derived distributions (Omarchy) on RTX Spark laptops encounter a firmware memory partition lock. Out of the box, the laptop’s UEFI/firmware carves out 96GB of the 128GB physical memory for the GPU using a hardware reservation register that only the Windows driver stack knows how to address or resize.
When booted directly into Linux with default factory settings, the Linux kernel only sees 28GB of total system memory—rendering the machine unable to load large models. To unlock the full memory pool for Linux, users must first boot into Windows, lower the Windows-side dedicated GPU memory reservation from 96GB down to 4GB, and reboot; only then does Linux detect 120GB of usable unified memory.
Why WSL2 Is the Mandatory Workaround Until Nvidia Updates Linux Drivers
Because WSL2 executes a real Linux kernel inside a lightweight hypervisor while delegating hardware power management and CUDA scheduling to the host Windows on ARM driver, it completely bypasses the encrypted power-table lockout. In WSL2, Linux-native inference engines like vLLM and SGLang sustain 65.1 tokens per second on GPT-OSS-120B—matching native Windows within 0.5% and even outpacing the desktop DGX Spark.
Use WSL2 Instead of Native Dual-Boot on RTX Spark Laptops : Run vLLM, SGLang, and PyTorch workloads inside WSL2 on Windows on ARM to retain the full 2.2 GHz GPU boost clock and 65.1 tok/s throughput on 120B models.Adjust the Windows GPU Memory Split Before Dual-Booting : If you must boot bare-metal Linux for kernel development, first lower the Windows GPU memory reservation from 96GB to 4GB so Linux can address 120GB of unified RAM.Choose Desktop DGX Spark for Pure Bare-Metal Linux Servers : If your deployment policy forbids Windows host OS layers, purchase the desktop DGX Spark (61.0 tok/s on native Linux) until Nvidia ports the Platform Controllers power logic to its open Linux kernel module.
Beat:
Topics:
# linux
# local-llm-inference
# nvidia
# rtx-spark
# wsl2
