RTX Spark Laptop on Linux: GPU Reportedly Runs at 780 MHz When Plugged In

A MindStudio write-up published November 1, 2026 by its director of product reports that native Linux on the MSI Prestige N16 Flip AI Plus laptop with Nvidia RTX Spark and 128 GB unified memory runs the GPU at roughly 780 to 790 MHz when plugged in, versus about 2.2 GHz on Windows, cutting GPT-OSS-120B generation to 13.9 tokens per second against 65.4 on Windows and 65.1 on WSL. The write-up attributes the gap to Nvidia Platform Controllers and Framework, whose encrypted power table it says only the Windows driver can decrypt, and reports that no software workaround restored full speed, with one attempt causing a kernel panic. No confirmation from Nvidia and no second independent test exists for the figures, which come solely from the author's own testing.

A MindStudio write-up claims that native Linux on the Nvidia RTX Spark laptop runs the GPU at roughly 780 to 790 MHz when the laptop is plugged in, against about 2.2 GHz on Windows. It says that cuts local LLM speed to about a quarter. The claim rests on one author's testing. We found no confirmation from Nvidia and no second independent test, so this post separates what is claimed from what is verified. It is a different product from the desktop covered in our DGX Spark posts https://explainx.ai/blog/nvidia-dgx-spark-ling-3-0-flash-moe-humming-fp8-lm-head-mtp-102-tokens-per-second-2026 . The Hacker News title mixed them up, and commenters noticed. TL;DR: claimed versus verified | Item | Claim | Status | |---|---|---| | Hardware | MSI Prestige N16 Flip AI Plus with RTX Spark and 128 GB unified memory | Product exists; MSI opened pre-orders October 8, 2026, per press coverage such as Guru3D https://www.guru3d.com/review/msi-prestige-n16-flip-ai-plus-c1xm-review-nvidia-rtx-spark-n1x-128gb/?ref=explainx | | Linux GPU clock on wall power | About 780 to 790 MHz vs about 2.2 GHz on Windows | One source, MindStudio https://www.mindstudio.ai/blog/linux-nvidia-rtx-spark-laptop?ref=explainx ; not verified | | GPT-OSS-120B speed | Windows 65.4 tok/s, WSL 65.1, DGX Spark 61, native Linux 13.9 | One source; not verified | | Cause | An encrypted power table read only by a Windows driver component | One source's explanation; Nvidia has not confirmed | | On battery | Linux comparable to Windows | One source; not verified | | Fix | Nvidia would need to port the power logic to its Linux driver | Opinion of the author | What MindStudio reports The article, published November 1, 2026 by MindStudio's director of product, describes dual-booting Linux on the laptop. The key numbers it gives: - GPU clock when plugged in: about 780 to 790 MHz on Linux, about 2.2 GHz on Windows. - GPT-OSS-120B generation: 65.4 tokens per second on Windows, 65.1 on WSL, 61 on a DGX Spark and 13.9 on native Linux. - vLLM with 8 concurrent clients: 127 tokens per second on WSL against 39.5 on native Linux. - A longer benchmark that took about 8 minutes on Windows took about 37 minutes on Linux. - LM Studio: 61 to 67 tokens per second on Linux on battery, about 15 plugged in. - Temperatures stayed at 40 to 47 degrees Celsius, which the author uses to argue this is power management and not heat. On cause, the author blames the Nvidia Platform Controllers and Framework, a component that splits the power budget between CPU and GPU. The write-up says its power table is encrypted and only the Windows driver can decrypt it. It says no software workaround restored full speed, and that one attempt caused a kernel panic. The article cites no outside source. It says its findings come from its own tests. We cannot reproduce them. The memory claims in the same write-up The article makes a second set of claims about memory. It says the laptop has 128 GB of unified memory with 96 GB reserved for the GPU by default, and that Linux at first saw only 28 GB. It reports that cutting the Windows-side reservation to 4 GB exposed about 120 GB. It also says WSL reserves 14 GB by default. Unified memory means the CPU and GPU share one pool instead of keeping separate RAM and video memory. That is why a 120-billion-parameter model can load at all on a laptop. The split between the two sides is a setting, and a wrong split can make a large model fail to load or run slowly. These details are plausible, but they come from the same single source, so they carry the same caveat. The article also reports that disabling the power controller on Windows dropped throughput from about 83 to 31 tokens per second. If correct, that supports the idea that the controller matters to speed on this chip. It still does not show that Linux cannot read the power table; it shows only that the controller has an effect on Windows. RTX Spark versus DGX Spark: why the names confuse people Both products use a similar Nvidia chip family with Arm CPU cores and Blackwell GPU cores. The DGX Spark is a small desktop box aimed at developers and ships with a Linux-based software stack. The RTX Spark is the name for the chip in Windows on Arm laptops and mini PCs from partners such as MSI. In the June 2026 forum post, a user noted the DGX Spark draws about 200 W at rated performance, while the RTX systems are rated at roughly 80 W for laptops and 100 W for desktops. The user asked whether that means lower clocks. This is a user's question, not an Nvidia answer, but it explains why power budgets are central to how these laptops perform. A laptop that can use its whole power budget on Windows and a smaller one on Linux would show up exactly as this write-up describes. For the full picture of the desktop line, read our earlier DGX Spark tests https://explainx.ai/blog/nvidia-dgx-spark-ling-3-0-flash-moe-humming-fp8-lm-head-mtp-102-tokens-per-second-2026 . The numbers there came from a different machine. What we could and could not verify Verified: the laptop exists and ships with 128 GB of memory shared between CPU and GPU. Retail and press listings describe Windows 11 on Arm. A June 2026 thread on the Nvidia developer forum https://forums.developer.nvidia.com/t/two-questions-for-nvidia-about-the-rtx-systems/372142?ref=explainx shows a user asking whether RTX Spark systems would run lower clocks than the DGX Spark and whether Linux could be installed. No Nvidia reply appears in the captured page. Not verified: the clock numbers, the token speeds, the encryption explanation, and the battery behaviour. We found no Nvidia release note, kernel bug report or second benchmark that matches. A search turned up no source stating that Linux is officially supported on RTX Spark laptops. Plausible but unproven: laptop GPUs on Linux have a history of power-limit bugs. Nvidia forum users have reported laptop GPUs stuck at low power states on wall power on older chips, and some fixes involved restarting the nvidia-powerd service. That shows the class of problem is real. It does not show that this specific claim is right. What Hacker News readers said The Hacker News thread https://news.ycombinator.com/item?id=50362905&ref=explainx had about 30 points and under ten comments when we read it. - User hypfer and user schaefer pointed out that the title said DGX Spark while the article is about the RTX Spark laptop, and that the Linux complaint makes sense only for the laptop. - User rich sasha worried that a locked power table could become a paywall or gatekeeping tool https://news.ycombinator.com/item?id=50363253&ref=explainx for running open models. That is the commenter's speculation, not something the article claims. - User fxwin complained that the article's subheading reads like generated text https://news.ycombinator.com/item?id=50363331&ref=explainx . That is a style opinion, but it is another reason to want a second source. Why it matters if true Local inference is one of the main reasons people buy 128 GB unified memory machines. Our local AI diary coverage https://explainx.ai/blog/verge-local-ai-diary-hermes-agent-qwen-mac-studio-lessons-2026 and the Qwen3 FP4 test on RTX hardware https://explainx.ai/blog/nvidia-qwen3-8-flash-next-fp4-local-96gib-nvfp4-rtx-pro-6000-dgx-spark-2026 show how many builders run models on Linux. If a Linux user got a quarter of the speed, the practical answer would be Windows or WSL for now. It also ties into the supply picture. See our look at the AI chip supply chain https://explainx.ai/blog/ai-chip-supply-chain-tsmc-record-samsung-phone-cut-globalfoundries-interposers-october-2026 , where memory and packaging decide how many of these machines ship. How you can check it yourself If you own an RTX Spark laptop, you can test the claim in a few steps: 1. Run the same model and prompt on Windows or WSL and on native Linux. 2. Log GPU clock and power with nvidia-smi during the run, on wall power and on battery. 3. Note the driver version, firmware and power profile. 4. Report the numbers to the Nvidia developer forum with the exact model name. Share any result with Nvidia and the community. A second test would turn this from a claim into a bug report. What to do now A single test can be wrong in many ways: a bad power profile, an old kernel, a missing firmware step, or a driver build that was never meant for this chip. The author also says one Linux distribution worked after driver fixes while another needed workarounds, so the setup was not clean. Keep that in mind before you repeat the 13.9 tokens per second figure as fact. - If you need Linux for local models, wait for Nvidia or the laptop maker to state Linux support in writing. - If you already use WSL, the write-up says speed matches Windows, which the author reports as 65.1 versus 65.4 tokens per second. - Do not buy based on this single post, in either direction. Frequently asked questions Does this affect the DGX Spark? The write-up lists the DGX Spark at 61 tokens per second as a reference point. It does not claim the DGX Spark has the problem. Could it be a bug rather than a deliberate lock? The author describes an encrypted table, but the cause is the author's reading. A driver bug is as likely an explanation until Nvidia says otherwise. Will we update this post? Yes, if Nvidia, MSI or another tester publishes data. Related reading Figures are as reported by MindStudio on November 1, 2026 and are not independently verified.