TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
A MindStudio report measured GPT-OSS-120B inference at 13.9 tokens per second on native Linux on an MSI Prestige RTX Spark laptop, compared with 65.4 on Windows and 65.1 in WSL. The report attributes the gap to Linux lacking access to a Windows-only power-management component and its encrypted power table; Nvidia has not provided a fix in the supplied source material.
Native Linux inference on an MSI Prestige RTX Spark laptop reached 13.9 tokens per second in a MindStudio benchmark, compared with 65.4 on Windows and 65.1 in Windows Subsystem for Linux (WSL). The report attributes the slowdown to the GPU running at about 780 MHz on Linux, rather than roughly 2.2 GHz on Windows, because Linux cannot access a Windows-only power-management component.
MindStudio tested the laptop with GPT-OSS-120B, a 120-billion-parameter model. It reported throughput of 65.4 tokens per second on Windows, 65.1 in WSL and 61 on Nvidia’s DGX Spark desktop. Native Linux on the laptop managed 13.9 tokens per second. These are results from the report’s test setup, not a general benchmark of every RTX Spark laptop or Linux configuration.
A separate test using vLLM with eight concurrent clients also showed a gap: the report measured 127 tokens per second in WSL and 39.5 on native Linux. In a longer run, Windows completed the task in about eight minutes while the GPU stayed near 2.2 GHz; Linux took 37 minutes, with clocks around 790 MHz. MindStudio said temperatures remained between 40 and 47°C, pointing away from thermal throttling in that test.
The report says Linux could boot and run a desktop after hardware and configuration work, including GPU acceleration through Nvidia’s open kernel driver. It also describes initial installation and hardware-support problems, and says the laptop’s memory split must be adjusted in Windows before Linux can access most of its unified memory. These results concern the MSI Prestige configuration described by MindStudio; the source does not establish that all models behave identically.
Linux AI Runs Far Below Windows
The performance difference matters to developers choosing an RTX Spark laptop for local AI inference. In the reported model test, native Linux throughput was about one-fifth of the Windows result, while WSL delivered nearly the same rate as Windows. That makes the choice of operating environment consequential for workloads that depend on sustained GPU performance.
Many AI development tools, including vLLM and SGLang, are built with Linux workflows in mind. The report suggests users may be able to run those tools through WSL without taking the same performance hit, but that is not the same as having a fully native Linux installation. For buyers, the gap also complicates the promise of using a high-memory, GPU-focused laptop as a portable development system independent of Windows.
The benchmark does not show that Linux itself is inherently slower on this hardware. MindStudio’s explanation focuses on platform-specific power control: the GPU appears able to accelerate under Windows and WSL but is held to a lower clock under native Linux. That distinction matters because a driver or firmware change could alter results.
Products Worth Considering
Apple MacBook Air 13.6-inch with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
A Windows-First Laptop Platform
The RTX Spark laptops described in the report, including models from MSI, Asus and Dell, pair an Nvidia GPU with unified memory shared by the CPU and GPU. The report says these systems ship with Windows on ARM, while Nvidia’s DGX Spark desktop runs Nvidia’s Linux build. Developers interested in Linux may therefore need to install it themselves or use WSL.
On the MSI Prestige, MindStudio describes a non-plug-and-play Linux setup. Debian initially lacked working support for several laptop components in the tested configuration, and the installation required workarounds. The author later used a Linux distribution with Nvidia’s open kernel driver to get a working desktop and GPU acceleration. The laptop’s firmware reserved 96GB of its 128GB unified memory for the GPU by default; after changing the allocation in Windows, Linux could see more memory.
The central issue identified in the report is Nvidia’s Platform Controllers and Framework, a component that manages power delivery between CPU and GPU. MindStudio says its power table is encrypted and can be decrypted by the Windows driver, while Linux has no equivalent mechanism in the tested setup. The article presents this as the explanation for the low Linux clocks, rather than as a finding independently confirmed by Nvidia.
“Windows produced 65.4 tokens per second. WSL hit 65.1 tokens per second. Native Linux on the laptop managed only 13.9.”
— MindStudio report
Driver Support and Model Differences
The supplied source material does not include a response from Nvidia, an announced driver update or a confirmed schedule for Linux support. It also does not independently verify MindStudio’s explanation that encrypted firmware data is the root cause of the clock limit. The benchmark results and diagnosis should therefore be read as the report’s findings and interpretation.
The article’s headline refers to inference speed falling to 35%, but the supplied benchmark figures require care: 13.9 tokens per second is about 21% of the Windows result of 65.4 in the GPT-OSS-120B test. The source does not explain that headline percentage or identify a different comparison behind it. Its separate vLLM test reports 39.5 on native Linux versus 127 in WSL, a different workload and comparison.
The source text also begins to discuss how results change when the laptop is unplugged, but the passage is cut off before giving a complete result. The effect of battery power cannot be established from the supplied material. Results on other laptop models, firmware versions, Linux distributions and driver releases are also not specified.
Nvidia Linux Power Support
The next relevant development would be Nvidia’s response or a Linux driver update that supports the laptop’s platform-controller power management. MindStudio says a fix would require bringing that logic into Nvidia’s Linux driver, but the supplied report gives no commitment or release date.
Until a fix is documented and tested, users considering these laptops for Linux AI workloads should treat the figures as specific to the reported setup and verify performance on their own configuration. The report identifies WSL as a current workaround for retaining Windows-level GPU clocks while running Linux-oriented tools, though it does not resolve the limitations of a native Linux installation.
Key Questions
How fast was native Linux in the reported test?
MindStudio measured 13.9 tokens per second with GPT-OSS-120B on native Linux on the tested MSI Prestige laptop. Windows reached 65.4 and WSL reached 65.1 tokens per second in the same reported comparison.
What does MindStudio say causes the slowdown?
The report attributes it to a Windows-only power-management component and an encrypted power table that Linux cannot access in the tested setup. Nvidia has not confirmed that explanation in the supplied material.
Does WSL perform like Windows?
In MindStudio’s GPT-OSS-120B test, WSL recorded 65.1 tokens per second, close to Windows at 65.4. A separate vLLM test also favored WSL, but the reported figures apply to those workloads and configurations.
Has Nvidia announced a fix?
No fix, release date or Nvidia response appears in the supplied source material. The report says Linux support for the platform controller’s power-management logic would be needed.
Source: mindstudio.ai
