benchmarks
- Part 2 · 7 min read
Memory bandwidth decides the speed
Five office PCs, a mini PC and a graphics card suggest that, for dense models, memory bandwidth rather than drive speed or core count sets how fast a model talks, with one mixture-of-experts model as the exception.
- Part 4 · 8 min read
An accelerator only helps if it brings its own memory
A graphics card with its own memory runs models fast until one no longer fits, while the extra processors in my mini PC share the CPU's memory and add little I could find.