Skip to content

Notes

First-hand measurements run on hardware I own, published with the scripts and the raw dataset so the numbers can be recomputed rather than trusted.

7 posts

Posts in this cluster

Notes2026-09-09

The Second User Buys No Batching At All. The Fourth Buys a Third of What Theory Promises.

Ten rounds of four requests on a fanless 8GB M1, dispatched one at a time, two at a time, and four at a time. Four-way concurrency raises server throughput 39% and cuts every user's generation rate by 65%. The two-way case is stranger: its 9% gain is not batching, and generation actually gets 6% worse.

Analysis Brief →
Notes2026-09-04

I Measured the Wrong Thing Three Times — What Animation Actually Costs a Reading Page

Swapping a JavaScript scroll reveal for the CSS one saved 16 ms of script per page scroll. Stopping eight animations that repeat forever saved nothing measurable. And the harness that found this told me a confident, wrong answer three times before it told me a true one.

Analysis Brief →
Notes2026-08-27

The Benchmark Number Is a Burst — I Ran a Fanless M1 Flat Out for 30 Minutes

746 back-to-back generations across two 30-minute runs on a passively cooled MacBook Air. Both settle on the same sustained rate — 51.7 and 50.9 tok/s — while approaching it from opposite directions, and macOS reports no thermal pressure in either.

Analysis Brief →
Notes2026-08-27

I Tried to Measure What One LLM Query Costs in Energy. The Instrument Lied Twice.

A failed measurement, published in full. The Mac battery gauge refreshes once a minute against generations lasting 2.8 seconds, and where a long enough window exists to beat that, its integral disagrees with the pack's own charge counter by 2.05x. Per-query energy is not measurable on this hardware without root.

Analysis Brief →
Notes2026-08-26

A 2.5× Bigger Model Costs Only 1.55× the Speed — I Measured Why on an 8GB M1

Fifteen benchmark runs across three local models on one MacBook Air, with the quantization variable held still. The result isn't the tokens-per-second table — it's what the table reveals about which physical resource local inference is actually spending.

Analysis Brief →
Notes2026-08-16

The Real Ceiling for Local LLMs on an 8GB Mac Isn't RAM — It's Disk

A follow-up benchmark on the same 8GB Apple M1: a genuine Llama 3.2 3B measurement, then a real, unplanned finding — macOS's own swap file, not GPU memory, is what actually stopped this machine from testing a 7-8B model.

Analysis Brief →
Notes2026-07-27

How to Run a Local LLM: Hardware, Cost, and Tokens-Per-Second Benchmarks

Real tokens-per-second numbers measured on an actual 8GB Apple M1 via Ollama, plus cited community benchmarks showing what changes once you move to a bigger model or a dedicated GPU.

Analysis Brief →