SYLEN
AboutNewsConferenceMembershipDonate

Email updates

Conference, news, and membership updates by email.

Site

  • About
  • News
  • Membership
  • Waitlist
  • Donate

Conference

  • Conference 2027
  • Call for papers

Account

  • Create account
  • Membership details

SYLEN

  • Guidelines
  • Privacy
  • Terms

© 2026 Systems Leadership and Engineering Network. sylen.org.

Membership details →
Back to news
General SESource: josie.lolOctober 19, 2026

Custom Arch Linux ARM Port Unlocks 116 GiB Unified Memory and Native vLLM Execution on RTX Spark Hardware

Arch Linux package maintainer Josephine Pfeiffer engineered custom kernel overrides and ACPI patches to run native Arch Linux ARM on the ASUS ProArt P16 laptop. The bringup bypasses Windows memory segmentation to expose a unified 116.0 GiB host/CUDA pool, outperforming native Windows benchmarks in multi-threaded CPU tasks and sustaining containerized local FP4 LLM inference.

Custom Kernel Bringup and Driver Integration

Bringup on the ASUS ProArt P16 H7607BA-SE033W required patching upstream Arch Linux ARM build scripts combined with custom hardware overrides. Pfeiffer integrated NVIDIA kernel work originally targeting Ubuntu, running Linux kernel `7.0.0-1022-nvidia` alongside NVIDIA driver `615.78.08`. Establishing a functional virtual console required resolving ACPI override allocations and framebuffer mapping parameters derived directly from dumped DSDT/SSDT tables.

Peripheral support necessitated a patched `hid-asus` driver for full keyboard support, a dedicated machine driver for the audio subsystem, and locally extracted H7607BA.300 speaker firmware from ASUS installer packages. System installation was executed via an extended `archinstall 4.5` installer holding custom kernel, NVIDIA, power management, and firmware packages using `IgnorePkg` flags while allowing standard rolling userspace updates. The target environment drives display output at 3840 × 2400 (120 Hz) using NetworkManager, PipeWire, and rootless Podman integrated with NVIDIA Container Toolkit (CDI).

Memory Allocation and Host/CUDA Unified Pool

The primary architectural advantage of the Linux port is addressable memory layout efficiency. The physical 128 GB memory subsystem on the ProArt P16 is managed under Linux as a single shared host-and-device pool reporting 116.0 GiB available directly to host OS tasks and CUDA runtime allocations.

In contrast, Windows partitions physical memory into dedicated GPU, shared GPU, and CPU-only regions, requiring static driver slider modifications and system reboots to reallocate dedicated VRAM reservations. Frandroid's Windows baseline on identical 128 GB hardware achieved a maximum dedicated VRAM allocation of 96 GiB and a total CUDA addressable budget of 109.7 GiB (when leveraging shared memory allocations). The native Linux setup provides a continuous 116.0 GiB addressable memory boundary—yielding an additional 6.3 GiB for CUDA workloads while retaining dynamic host memory access without rebooting.

CPU Performance and Microbenchmarks

Synthetic CPU compute was benchmarked using native ARM64 7-Zip 26.04 under a fixed 32 MiB dictionary size across single-thread and 20-thread workloads on the 20-core N1X processor architecture:

  • Single-Thread: Recorded 6.18 MiB/s compression (7.221 GIPS) and 118.26 MiB/s decompression (10.779 GIPS). This matches published Windows ARM64 benchmarks on the HP OmniBook Ultra 16 (7.31 GIPS compression, 10.87 GIPS decompression) within a 0.8% to 1.2% delta.
  • 20-Thread Multi-Core: Achieved 141.77 MiB/s compression (165.754 GIPS) and 1,629.60 MiB/s decompression (148.471 GIPS). Under full thread saturation, the Linux setup outperformed native Windows on the HP OmniBook Ultra 16 by 5.0% on compression (165.754 vs. 157.9 GIPS) and 2.7% on decompression (148.471 vs. 144.5 GIPS). Total median execution (157.084 GIPS) also exceeded ComputerBase Windows 7-Zip 26.00 figures (155.774 GIPS) on 64 GB hardware.

Local Containerized LLM Execution with vLLM

Inference workloads evaluated NVIDIA’s `Qwen3.8-Flash-Next-FP4-Local` checkpoint running within native ARM64 rootless Podman containers using vLLM, B12X CUDA kernels, mixed quantization, and FP8 KV caching.

Starting from 110.9 GiB of free system memory, vLLM allocated 68.2 GiB for FP4 model weights and 21.1 GiB for the KV cache, establishing an ~89.3 GiB active hardware footprint. Context limits were bound to 32,768 tokens. Real-world task metrics demonstrated low TTFT (Time to First Token) and consistent generation rates:

  • Short Code Review (208 prompt tokens): 0.316 s TTFT, median end-to-end response time of 4.008 s, yielding a generation speed of 30.9 tokens/s. This matches Windows llama.cpp UD-IQ4_XS execution benchmarks (29.4–31.4 tokens/s).
  • Long Incident Context (9,775 prompt tokens): 4.532 s TTFT, processing the context and generating the target output in 5.768 s total (6.8 tokens/s overall pipeline rate across context and generation).
  • Structured Data Extraction (87 prompt tokens): 0.256 s TTFT, 2.035 s complete generation time.

Under a concurrent load of four parallel code review streams, the system achieved a combined generation throughput of 86.6 tokens/s with a median batch execution time of 5.314 s. A 10-minute sustained burn-in executing 150 continuous requests across three tasks demonstrated thermal stability, with median response latencies shifting by less than 1.6% between initial and terminal execution windows.

Read the original article at josie.lol.