Putting Linux on my RTX spark laptop because I don't have a Microsoft account

When rumors started circulating about a laptop with over 100 GiB of RAM, the writer began to imagine the possibilities - all the things they could build and optimise on it. They are a big proponent of open-weight models and use Hugging Face to publish their own GGUF quants, which have been downloaded over 160,000 times in the past month.

They were disappointed to find out that the official Windows support for this laptop was only available for Windows, and it made them less enthusiastic about the machine. However, they preferred using Arch Linux as their distribution and decided to set it up on their ASUS ProArt P16 H7607BA-SE033W laptop.

The writer used Arch Linux ARM and added hardware support on top of it. They incorporated NVIDIA's kernel work for Ubuntu, fixed ACPI override allocation and framebuffer mapping, and used the laptop's own ACPI tables and speaker firmware. They also patched the keyboard and audio drivers. This setup allowed them to run a bootable USB with their hardware support and Archinstall 4.5, which installed all the necessary hardware packages.

The display runs at 3840 × 2400 and 120 Hz, and they use NetworkManager, PipeWire, and rootless Podman with NVIDIA CDI.

To compare the performance of their Linux setup with the Windows version, they ran the 7-Zip benchmark using native ARM64 binaries. Their Linux setup achieved a score of 157.084 GIPS, while the Windows version scored 155.774 GIPS. The writer's Linux setup reported 116.0 GiB of system RAM and CUDA's total memory pool, while Windows divided the RAM into dedicated GPU, shared, and CPU-only regions.

Their Linux setup reported about 6.3 GiB more memory available for CPU work, which allowed them to run mixed quantization and FP8 KV cache for vLLM with B12X CUDA kernels in a native ARM64 container. The model loading and compilation took about 16 minutes, and the tasks ran three times after warmup, with known answers to check against.