Observed arrival · 2026-11-20
Local LLM speed, measured against the roofline
Boltbeam compares local LLM speeds across GPUs, models, quantizations, and engines, with measurements set against each GPU’s estimated roofline.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- People comparing GPUs for local LLM inference
- Worth noticing
- For this run, the page says no per-role timing or kernel capture was available.
Field notes
The page distinguishes stock llama.cpp results from results using modifications and tuned settings, and labels which setup is fastest for each listed GPU. For the featured run, it gives context bands from under 32k to 96–128k and describes paired runs, a 450 W power cap, and temperature gates. It also cautions that this run timed the whole step only, with no per-role split or kernel capture.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue