The Nvidia RTX Spark platform is a specialized hardware configuration designed for local AI development. It features a unique unified memory architecture—supporting up to 128GB of memory shared between the CPU and GPU—and utilizes the same GPU design as Nvidia's DGX Spark desktop units. While these laptops are marketed to developers who prefer Linux-first AI tools like vLLM and SGLang, the hardware is currently optimized for Windows on ARM, which includes specific driver support for Nvidia's CUDA stack.
Testing reveals a significant performance disparity between operating systems when the laptop is connected to wall power. On a machine running the GPT-OSS-120B model, native Windows and the Windows Subsystem for Linux (WSL) both achieved approximately 65 tokens per second. In contrast, native Linux installations on the same hardware dropped to 13.9 tokens per second. This slowdown is not a result of thermal throttling, as temperatures remained low (40-47°C), but rather a clock speed collapse. While the GPU maintains a boost of roughly 2.2 GHz on Windows, it becomes locked at approximately 780 MHz on Linux when plugged in.
This performance gap is attributed to a Windows-only component called the Nvidia Platform Controllers and Framework. This firmware and driver logic manages the shared power budget between the CPU and GPU, a necessity for laptops where the two chips share a power line. The firmware provides a table of power settings that is encrypted; while Windows possesses the keys to decrypt this table, Linux currently does not. Interestingly, the performance issue reverses when the laptop is unplugged: on battery power, Linux performance improves to roughly 61-67 tokens per second, nearly matching Windows levels. Until Nvidia ports the platform controller's power management logic into its Linux drivers, WSL remains the most practical middle ground, as it inherits the Windows driver stack's ability to maintain full GPU boost while still providing a Linux-compatible environment.
One notable limitation of this setup is memory management. The laptop's firmware reserves 96GB of unified memory for the GPU by default, a split that is currently only adjustable through Windows. Users attempting to run native Linux must first adjust this reservation in Windows to free up enough memory for the Linux environment to see more than a small fraction of the total capacity. While the source material confirms that the issue is a firmware and driver gap rather than a hardware defect, there is currently no confirmed timeline for when Nvidia might release a Linux-compatible power management driver for these units.
One specific benchmark using vLLM with eight concurrent clients further highlights the disparity: WSL reached 127 tokens per second, while native Linux managed only 39.5. In a longer test, Windows completed the workload in about 8 minutes while maintaining a 2.2 GHz clock speed, whereas Linux took 37 minutes with the GPU parked at 790 MHz. These figures underscore the impact of the missing power management logic on high-throughput AI workloads.
This image illustrates the performance gap described in the benchmarks, where native Linux experiences severe throttling compared to Windows and WSL.
Ultimately, the current state of the RTX Spark laptop creates a trade-off for developers. While native Linux offers a pure environment, the lack of access to the encrypted power management tables effectively cripples the hardware's performance when plugged in. For the time being, developers seeking the best balance of Linux-native tooling and full hardware performance are directed toward WSL, provided they manually increase the default 14GB memory allocation to accommodate larger models.