Talos Linux vs Ubuntu

The same QuietBox 2 and DGX Spark (alone and as a pair), reinstalled with Talos Linux and serving the same model builds as Kubernetes pods, against the Ubuntu runs on the main page. Each Talos test is paired with the identical Ubuntu test. Median change in throughput from Ubuntu: —.

By model

Change on Talos, by model

Each dot is one test (a workload at one user count); the bar is the model's median.

By use case

Change on Talos, by use case

Each dot is one model at one user count; the bar is the median for that system.

Every test

Change on Talos, every test

Percent change from Ubuntu to Talos. Columns are workloads, split by the number of concurrent users.

Reading these results

  • Same hardware, model builds, prompts and load on both. What changed with the OS: the kernel (Talos 1.14 runs 6.18 and Talos 1.15 runs 7.3, against Ubuntu's 7.0), the NVIDIA driver on the Spark (595 on both Talos versions, against 580 on Ubuntu), and the runtime: Kubernetes pods under containerd on Talos, tt serve and Docker on Ubuntu. The Spark ran the same vLLM image, by digest, everywhere.
  • Talos 1.15 (an alpha release, 1.15.0-alpha.2) has only been run on the DGX Spark so far. On 1.15 the GPU reaches the pod through CDI and the NVIDIA container toolkit's NRI plugin instead of a runtime class.
  • The one large drop on Talos 1.15, gemma-4-31b at 32 users (about −24%), is a single run: every request finished, but each token took longer. Its other tests sit with the rest.
  • The 2x DGX Spark rows run tensor parallel 2 across both Sparks over the QSFP cable, with NCCL using RoCE on both rails on Talos and on Ubuntu. The Talos nodes boot with pci=pcie_bus_safe: without it the ConnectX-7 runs with a 128-byte PCIe payload and one rail tops out near 97 Gb/s instead of about 110 Gb/s. The vLLM image's CUDA forward-compat libraries are loaded ahead of the host's 595 driver libraries; without them NCCL crashes at startup on this image. Two models so far, gpt-oss-120b and Qwen3.5-122B-A10B; Llama 3.3 70B has not been run on the pair under Talos.
  • The QuietBox 2 rows use the builds that ran on Ubuntu: the v1 bundles, not the fused-decode rebuilds the main page shows, which only ran on Talos.
  • Most QuietBox 2 differences are small, but under load several models reached their first token later on Talos, which cuts total throughput at 16 and 32 users. Single runs of four requests (the 1-user tests) move a few percent from run to run on either OS, so read single dots with that in mind and lean on the medians.

Method. Every run sent random prompts with a fixed output length and greedy sampling (temperature 0) through vllm bench serve. Each Talos test is matched to the newest Ubuntu run of the same model build, workload, prompt length and user count; the change is (Talos − Ubuntu) / Ubuntu.

Data and code: talos.csv and the benchmark harness at github.com/rothgar/personal-ai-hardware. Built 2026-11-21.