Skip to content
Open WebUI to FastFlowLM to the XDNA2 NPU on IPv6 loopback: 21 W answering, 232,000 tokens in one prompt, three years from AMD's keynote to Linux

Running An LLM On The Ryzen AI NPU, In A Container

The NPU in a Ryzen AI laptop is 50 TOPS that sits idle on most Linux installs. This walks through building AMD’s amdxdna driver and keeping only the kernel module on the host, building FastFlowLM with AMD’s XRT runtime into a rootless podman container matched to that driver, binding everything to IPv6 loopback and starting it at boot with Quadlet. Then what it does: nine models benchmarked, a 232,000-token context test, Whisper speech-to-text at ten times real time, Open WebUI and the 7,000 hidden tokens it was adding to every message, and the speed and energy of the NPU measured against llama.cpp on the Radeon 860M iGPU. Last, the timeline: three years from AMD announcing Ryzen AI to a language model on its NPU under Linux, delivered by a project AMD did not start, while the effort went to Copilot+ on the Fisher-Price OS (Windows) instead of the data centre and the Linux tooling AI professionals actually use.

13th November 2026 · 24 min · 5853 words · Damien Dye

Sorting TSC When Your Firmware Is Broken

On every boot for four months my ThinkPad E595 handed Linux one CPU core three seconds behind the other seven, and the kernel dropped the TSC for HPET. What the warning means, why an AMD Zen machine has no way to repair it from inside Linux, what 40 boots of journal say about the firmware, why tsc=reliable is the wrong answer, and how tscsync measures and moves the counters in UEFI before the kernel ever sees them, including the one core it marked BAD that Linux accepted anyway. Updated for BIOS 1.30: it does not fix the offset, and the kernel’s clocksource watchdog dropped the TSC on two boots even after tscsync synced the cores. With tsc=nowatchdog it held, and a five-minute check against the real-time clock, source included, found no warps and no drift. Last, what the fallback does on any KVM hypervisor, Proxmox or not: the master clock goes, guests lose their fast clock read, and on this laptop a guest clock read came out seven times slower.

11th October 2026 · 23 min · 5100 words · Damien Dye
An NVMe drive entering a tri-mode adapter and leaving as a SCSI disk with a queue depth of 128

Tri-Mode Adapters Buy Flexibility With Your NVMe Queues

A tri-mode adapter lets any bay take SAS, SATA or NVMe — real flexibility, and the reason U.3 backplanes exist. What it does not advertise is that your NVMe drives stop being NVMe drives: they arrive in Linux as SCSI disks on mpt3sas, queue depth 128, sharing one tag pool and one x8 uplink. Two Gen4 drives saturate the card. A single Gen5 drive is already past it. Whether that trade still makes sense in 2026, and why the universal bays belong to the backplane rather than the controller.

13th August 2026 · 25 min · 5202 words · Damien Dye
A small aperture sliding over a large framebuffer, beside the whole framebuffer mapped at once

PCIe Resizable BAR and Modern GPUs — Intel Arc, NVIDIA and AMD

Resizable BAR lets the CPU map a GPU’s whole framebuffer instead of peering at it through a 256MB window. Intel calls it required for Arc, NVIDIA enables it per game, AMD sells it as Smart Access Memory — and for AI work it changes the transfer, not the maths.

7th August 2026 · 11 min · 2295 words · Damien Dye