Encrypted Power Firmware Throttles Nvidia RTX Spark Laptops to 780 MHz Under Native Linux
Dual-booting native Linux on Nvidia RTX Spark laptops degrades inference throughput by nearly 80% on wall power due to an encrypted firmware power table that only the Windows driver can read. Without the Nvidia Platform Controllers and Framework driver present on Linux, the GPU clock drops to a fail-safe ~780 MHz, leaving WSL as the only viable environment for high-performance workloads.
Unified Architecture and Memory Pre-Allocation Constraints
Nvidia's RTX Spark platform—featured in mobile workstations like the MSI Prestige N16 Flip AI Plus, Asus ProArt, and Dell XPS—pairs ARM host processors with CUDA-capable silicon over up to 128GB of unified memory. While architecturally close to the desktop DGX Spark, these mobile implementations introduce specific firmware-level constraints when booting native Linux.
By default, the system firmware reserves 96GB of unified memory exclusively for the GPU. Because the interface to modify this unified RAM allocation is exposed only within Windows, booting a native Linux kernel without prior adjustment leaves the OS with access to just 28GB of system memory. Lowering the GPU reservation to 4GB via Windows allows Linux to map 120GB of usable system RAM. Standard kernels also lack the peripheral and display pipe drivers for these ARM-based boards, requiring distributions with bundled open-source Nvidia kernel modules, such as Omarchy, to drive internal 4K 120Hz displays and onboard I/O.
Clock Speed Throttling and Benchmark Discrepancies
Once initialized with GPU acceleration through Nvidia's open kernel driver, native Linux exhibits a severe power-state performance anomaly. Running GPT-OSS-120B on AC power, native Linux drops to 13.9 tokens per second, with GPU clock speeds clamped to ~780–790 MHz. Counterintuitively, disconnecting wall power restores GPU clocks to ~1,800 MHz, yielding 61 to 67 tokens per second on battery power.
On AC power, the clock lock limits throughput across all optimized inference engines:
- GPT-OSS-120B (Single Client): Windows 11 achieves 65.4 tok/s, WSL reaches 65.1 tok/s, the native Linux DGX Spark desktop reaches 61 tok/s, while native laptop Linux hits 13.9 tok/s.
- vLLM (8 Concurrent Clients): WSL achieves 127 tok/s, whereas native Linux hits 39.5 tok/s.
- Execution Latency: A standard benchmark run taking 8 minutes on Windows at ~2.2 GHz requires 37 minutes on native Linux at ~790 MHz.
With operating temperatures holding between 40°C and 47°C during execution, thermal throttling is ruled out. The GPU core is artificially constrained by low power limits.
Root Cause: Unreadable Firmware Power Tables
The bottleneck stems from the power delivery architecture required for hybrid laptop designs. Unlike the desktop DGX Spark, which feeds the GPU over dedicated power lines, mobile RTX Spark systems rely on a Windows component called the Nvidia Platform Controllers and Framework. This module dynamically arbitrates thermal and electrical budgets between the CPU and GPU.
The underlying system firmware passes an encrypted power table to this controller. The proprietary Windows driver contains the decryption keys and management logic required to read the table and adjust dynamic power states. Nvidia’s open-source Linux kernel driver currently lacks this decryption path. When AC power is connected without an active platform controller driver, the GPU drops to a hardcoded 780 MHz fail-safe state.
Disabling the Nvidia Platform Controller inside Windows reproduces the exact behavior observed in Linux: token throughput falls from ~83 to 31 tok/s, and clocks lock to 786 MHz. Manual attempts to bypass or force driver power states under Linux without valid decrypted tables result in immediate kernel panics.
Production Workaround via WSL
Until Nvidia ports its platform controller power management logic to the Linux driver stack, native dual-boot configurations remain unsuitable for production AI workloads.
For engineers running Linux-native toolchains like vLLM or SGLang, Windows Subsystem for Linux (WSL) serves as the primary workaround. Because WSL runs as a hypervisor client over the Windows host driver stack, it inherits full GPU boost behaviors up to ~2.2 GHz regardless of power source. To handle large models, engineers must manually override WSL's default 14GB memory ceiling inside `.wslconfig` to fully utilize the unified 128GB address space.