# pkillsleep

> Personal site of a French hobbyist who runs AI locally on second-hand hardware (RTX 3090 + RTX 3080 Ti, 80 GB of RAM) and self-hosts his services. Every experiment gives measurements made by the author, with their conditions and limits. Raw data is in JSON. The site is written in French first; these English pages are translations reviewed by the author.

French version: https://pkillsleep.dev/llms.txt

Every page also exists in Markdown: add `.md` to the URL (e.g. `/en/labo/undervolt-3090-3080ti.md`).
All measurements: https://pkillsleep.dev/data/runs.json (format described at https://pkillsleep.dev/en/data/).

## Experiments

- [two coding agents on the same box, from 44 to 66 tok/s each in one evening](https://pkillsleep.dev/en/labo/deux-agents-44-66-tok-s.md): same machine, same model, one night of tuning: when two agents generate at the same time, each one goes from 43.8 to 66.4 tok/s, and four settings did most of the work (2026-10-20)
- [glm 5.3 flash (321B) on a 3090 + a 3080 ti: from 26 to 31.7 tok/s, and why i'm still sticking with qwen](https://pkillsleep.dev/en/labo/glm-53-maya-3090-3080ti.md): two maya options give 22% more decode to a 321B model that spills way over into ram, but on a real agent task it stays clearly slower than qwen 3.8 flash next on the same box (2026-10-20)
- [a 4B model holds 256K of context on a single 12gb 3080 ti, and finds 40 codes out of 40](https://pkillsleep.dev/en/labo/spark-4b-256k-3080ti.md): spark-x2.6-4b in q8 with llama.cpp: 136 tok/s empty, still 35 tok/s with 254K tokens of context, and a 12gb card that beats the 3090 because i had capped it less (2026-10-20)
- [i undervolted my 3090 and my 3080 ti on linux, -94 W for almost nothing](https://pkillsleep.dev/en/labo/undervolt-3090-3080ti.md): two coding agents working at the same time, 452 W → 358 W across both cards, the fans go from 91% to 48%, and throughput loses 7% (2026-10-20)
- [i fit 400K of context per agent on 80gb of ram, and it still finds a code hidden at 92% of the prompt](https://pkillsleep.dev/en/labo/400k-contexte-80go.md): two slots of 409,600 tokens fit on 80gb of ram, and a 373,723 token prompt with three codes hidden inside gets 3 right answers out of 3, read in 195 seconds (2026-10-19)
- [i asked for 512K of context × 4 slots on 80gb of ram, and the whole machine froze](https://pkillsleep.dev/en/labo/512k-machine-figee.md): at first i blamed a dead wifi dongle, but it was the ram: the session cache at 512K needed ~23gb more on top of a 40gb model that was already taking up the space, and a simple memory cap would have avoided it all (2026-10-19)
- [strata 0.2.0 on my rtx 30 cards: +12 to +17% decode, but my agents don't finish any faster](https://pkillsleep.dev/en/labo/strata-020-ampere.md): a simple engine update takes solo decode from 91.7 to 107.6 tok/s on a 3090 + 3080 Ti, and yet the total time of my agents doesn't move, because they spend their life reading prompts (2026-10-19)

## Guides

- [your first local llm with llama.cpp, on a 12gb card](https://pkillsleep.dev/en/guides/llama-cpp-premier-llm-local.md): compile llama.cpp for your nvidia card and run a recent model with 256K of context on 12gb, while understanding what each option does
- [running glm 5.3 flash (321B) on two consumer cards with maya](https://pkillsleep.dev/en/guides/maya-glm-53.md): how to install project maya, pick the right version of the model when your cards have 24gb or less, and the two options that got me 22% more speed, with what they cost
- [qwen 3.8 flash next on strata: a 125B at more than 110 tok/s on a 3090 + a 3080 ti](https://pkillsleep.dev/en/guides/strata-deux-gpu.md): how i install strata and run qwen 3.8 flash next on two consumer cards at more than 110 tok/s solo, then how i share it between two agents at the same time
- [undervolting a 3090 or a 3080 ti on linux, without losing performance](https://pkillsleep.dev/en/guides/undervolt-3090-3080ti-linux.md): the step by step i use on my box: lock the clock, shift the voltage/frequency curve with NVML, check under real load, then make it automatic at boot

## Optional
- [About](https://pkillsleep.dev/en/about/): who writes, how measurements are made
- [Data format](https://pkillsleep.dev/en/data/): the fields of runs.json