A personal lab

Building, measuring, and restoring computing systems.

I’m Paul D. Martin, Ph.D., and Technicomp Labs is my independent computing laboratory. This site is where I publish the open-source projects I release and my applied research, and write about restoring and upgrading vintage computers, particularly high-end systems, to their maximum performance. It is a personal, noncommercial project that follows my interests in computer history and vintage computing, in systems and virtualization, in open source infrastructure, and in technology education, and it is not client work.

Blog → Projects → About me →

Benchtop Linux

Details →

Technicomp Benchtop Linux · Alpha

The operating system for the technical workbench.

Benchtop Linux is a stabilized workstation rolling release derived from openSUSE Aeon and built on Tumbleweed. Most of the system stays current with Tumbleweed, while the kernel and desktop move more cautiously to reduce disruptive workstation regressions.

The operating system is delivered as a complete, immutable image, while applications and additional tools live in the user's profile through Flatpak, AppImage, and Homebrew rather than being layered into the system.

About Benchtop Linux → Source → Build Service →

Release: Stabilized workstation rolling release
Base: openSUSE Aeon / Tumbleweed
Kernel: LTS by default · Current for hardware enablement
Desktop: Blueprint Environment · previous supported GNOME release
System: Immutable image, transactional updates, snapshots
Apps: Flatpak, AppImage, and Homebrew, per user
Development: Major language toolchains included
Latency: Low interactive latency · real-time audio ready
Hardware: x86 and ARM
Status: Alpha images on OBS

Applied research

Lab update · October 24, 2026

Large local models at around 30 tokens per second.

I’m now running GLM-5.3-Flash Q4_K_XL and DeepSeek-V4.1-Flash Q8_K_XL at around 30 tokens per second with Project Maya on Galactus: four AMD Radeon Pro V620 GPUs, an EPYC 7713, and 2 TiB of RAM. The October 23 tests measured 31.6 tokens/s for GLM with default speculation and 29.3 tokens/s for DeepSeek with DSpark drafting enabled.

The following morning’s DeepSeek dashboard adds a view of ongoing use: 90 requests and 86,075 output tokens at a reported aggregate 28.0 tokens/s. The last request generated 422 tokens at 31.6 tokens/s. These mixed requests provide a useful follow-up to the controlled measurements, with the configurations and measurement limits recorded in the notebook.

Project Maya v1.1.24 serving DeepSeek-V4.1-Flash on four Radeon Pro V620 GPUs, with 31.6 tokens per second for the last request — DeepSeek on Galactus, October 24. The large number is the last request’s generation rate; the session aggregate is shown at the bottom. The dashboard is idle at capture time. Select the image to see it full size.

Read the measurements and source logs →

LLM Performance Engineering Notebook

Finding the real speed limit of local inference.

The notebook records my experiments on local inference, beginning with large Mixture-of-Experts models. I use hardware measurements and performance estimates to identify limits, then test configuration and code changes to understand where time goes. It includes a llama.cpp scheduler patch that raised prefill throughput 13.7% in the July comparison, per-model results, raw logs, and the hypotheses that testing did not support.

Read the notebook →

Galactus, the lab's inference server: four GPUs and an EPYC processor in a compact chassis

Hardware and history

Electronics workbench with parts organizers, an Apple IIe, and an open-frame test bench

The Lab

Bench, rack, and instruments

The bench, instruments, and servers I use to measure performance, diagnose hardware, and restore older machines.

The Collection

Every era at warp speed

The pinnacle of computing from every decade, restored and upgraded to the limits of its era.