RTX Spark on Linux Runs 61 Tokens a Second on Battery and 15 on Wall Power

The headline figure circulating on Hacker News, that Nvidia's RTX Spark laptops run inference at 35% of normal speed on Linux, does not appear in any of the three write-ups of the test that this article could fetch. The nearest number is arithmetic: the GPU clock reported on wall power, about 780 MHz, is roughly 35% of the 2.2 GHz it reached under Windows. Throughput fell further than that.

The underlying result is real but narrow. Neoteo reported on November 1 that Alex Ziskind, in a video published October 31, ran LM Studio on native Linux on a single MSI Prestige N16 Flip AI Plus and measured 61 tokens per second unplugged and 15 plugged in. A laptop that is faster on battery than on its charger is a firmware or driver fault, not a design choice.

What the test measured

The longer write-up, published by MindStudio on November 1, gives the GPT-OSS-120B comparison. The unit is generated tokens per second, one stream, and the article does not state the quantisation or context length, so the figures cannot be compared with anyone else's runs.

1. Windows65.4 tokens/s
2. WSL65.1 tokens/s
3. DGX Spark (native Linux)61 tokens/s
4. RTX Spark laptop, native Linux13.9 tokens/s

Source: MindStudio, Linux on the Nvidia RTX Spark Laptop, accessed 2026-11-04

The number that matters here is not the 13.9 but the 61. The DGX Spark desktop uses the same GB10-class silicon, and it runs native Linux at 61 tokens per second in this comparison. So the chip is not slow on Linux. The laptop is slow on Linux, on wall power, in this one configuration.

MindStudio also reports a longer run in which Windows finished in about 8 minutes at roughly 2.2 GHz while Linux took about 37 minutes at roughly 790 MHz, with temperatures between 40 and 47 °C. Cool and slow points to a limit set in software, not a thermal throttle.

The suspected cause is a Windows-only controller

Ziskind attributed the drop to "missing support for NVIDIA Platform Controllers and Framework in Linux", according to Neoteo, which presents that as his explanation and not a verified finding. MindStudio adds that the controller's power table is encrypted and that only the Windows driver can decrypt it. It says one software attempt to enable it triggered a kernel panic.

MindStudio also reports that disabling the controller in Windows dropped throughput from about 83 to about 31 tokens per second and pinned clocks at 786 MHz, and that re-enabling it restored normal speed. That is the strongest evidence in the piece, because it reproduces the Linux symptom on the same hardware without changing the operating system. It is also a single-source claim, and MindStudio does not name who ran it.

Condition (LM Studio, native Linux)	Tokens per second	GPU clock (MindStudio)
Unplugged	61	about 1,800 MHz
Plugged in	15	about 780 MHz

The unplugged clock and both throughput figures come from different outlets: Neoteo gives 61 and 15 and no clocks; MindStudio gives the clocks and a 61 to 67 range for the unplugged run. The two agree on the direction and roughly on the size.

Why a 128 GB laptop makes this a procurement question

The appeal of RTX Spark for local models is memory. ComputingForGeeks' spec roundup, checked against Nvidia, Microsoft and OEM pages on October 3, lists up to 128 GB of LPDDR5X shared between CPU and GPU and a 45 to 80 W laptop power envelope. Its own estimate is that 128 GB fits a 120B mixture-of-experts model.

On Linux that memory is not freely available either. MindStudio says the firmware reserves 96 GB for the GPU by default and only Windows can change the split. Setting the Windows-side reservation to 4 GB let Linux see about 120 GB, up from 28 GB. A buyer who wants Linux has to install Windows first.

Nvidia has not announced Linux support for the laptops. ComputingForGeeks recorded on October 3 that the spec sheet lists only Windows 11 and that Nvidia declined to comment to XDA. Community enablement exists: a GitHub repository by the user johannhartmann documents an ASUS ProArt P16 H7607BA on Omarchy with kernel 7.2.8-23 and driver 615.78.08. Its README says nothing about a wall-power throttle, and it records that the installer needs AC power.

What would change this read

Three things limit what can be concluded. The test covers one laptop and one workload. The Linux slowdown has not, in anything fetched here, been reproduced on the ASUS, Dell or other models. And Nvidia has said nothing in the coverage about the encrypted table or a driver fix.

The finding would be falsified by a second owner running the same LM Studio test on a different RTX Spark laptop and getting 61 tokens per second plugged in. It would be confirmed by Nvidia shipping Linux support for the platform controller; the open GPU kernel modules 615.71.09 that ComputingForGeeks examined list GB10 but no N1X device IDs, so that work has not started in public.

Until then, the practical benchmark for a Linux buyer is the number on wall power, and the only published one is 15 tokens per second. For local models on this hardware, the 27B-class open weights that people are running on laptops and Nvidia's own Nemotron releases tuned for RTX Spark are targeted at Windows first.