hn.today

Nvidia DGX Spark encrypted firmware slows inference speed to 35% on Linux

mindstudio.ai422 points192 comments
Screenshot of Nvidia DGX Spark encrypted firmware slows inference speed to 35% on Linux

Linux can be installed on Nvidia RTX Spark laptops such as the MSI Prestige N16 Flip AI Plus, with Nvidia’s open driver enabling CUDA, a working desktop, and high-resolution display output after hardware-specific setup. Installation is not straightforward: some distributions have display or driver problems, Debian may require wired networking and manual support, and Windows must be used to change the laptop’s default memory allocation. The machine reserves 96GB of its 128GB unified memory for the GPU, leaving Linux access to only 28GB until that setting is reduced in Windows.

The main obstacle is a severe, unusual performance drop when Linux is running on wall power. In GPT-OSS-120B tests, native Linux delivered 13.9 tokens per second, compared with 65.4 on Windows, 65.1 in WSL, and 61 on Nvidia’s DGX Spark. The GPU runs at about 780MHz under Linux when plugged in, rather than roughly 2.2GHz on Windows; temperatures remain low, ruling out overheating. Tests point to Nvidia’s Windows-only Platform Controllers and Framework, which manages CPU-GPU power sharing using an encrypted firmware table Linux cannot access. On battery, Linux performance approaches Windows levels. Until Nvidia supplies equivalent Linux power-management support, WSL is the practical option for Linux-based AI tools, though its memory allocation must also be increased for large models.

Read on mindstudio.ai192 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Hardware

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.