Nvidia firmware caps RTX Spark inference to 1/5 on Linux
An encrypted, Windows-only power table pins the GPU near 780MHz and Nvidia won't share keys — plus a live GitHub Actions outage and Opus 4.8 fast mode's exit.

Copy markdown
RTX Spark inference craters to a fifth on native Linux
A MindStudio GPT-OSS-120B run clocked 13.9 tok/s on native Linux versus 65.4 on Windows and 65.1 under WSL — an 8-minute job that took 37 minutes on Linux. The cause is an encrypted, Windows-only NPCF power table that pins GPU clocks at ~780MHz on wall power instead of the 2.0–2.2GHz Windows gets.
Nvidia won't hand over the keys — but WSL dodges it
Asked for Jetson-style owner keys to patch the firmware, Nvidia replied "We will not be providing keys for installing custom FW," with no Linux fix timeline. For now WSL matches Windows speed (set memory=96GB in .wslconfig), running on battery recovers Linux to ~61 tok/s, and the desktop DGX Spark on DGX OS is unaffected.
GitHub Actions and Pages are degraded right now
GitHub has been investigating an Actions and Pages incident since 16:05 UTC with no resolution posted — the second degradation today, after a morning outage hit Git Operations, Pull Requests and Issues until 11:50 UTC. If your CI is red or a Pages deploy is stuck, it's not your pipeline.
Opus 4.8 fast mode is on the clock in Claude Code
Claude Code 2.1.316 deprecated fast mode for Opus 4.8 on Oct 30 and says it will be removed Nov 30 — budget for the slower default now. The newer 2.1.317 also adds a check that asks before a web or shell tool call sends your account email over the network, plus CLAUDE_CODE_SUBAGENT_EFFORT to set subagent effort.