Guides
Data centre power, in plain words.
Six plain guides to the questions people ask about data centre energy, GPU power and capacity. Each one answers the question first, then says where ATHLAZ fits.
In reading order
- Guide 01
Energy efficiency in data centres
Data centre energy efficiency is useful work per unit of electricity. It rises when cooling and conversion take less, or the equipment draws less for the same work.
4 minute read - Guide 02
PUE: what it measures and what it misses
PUE is a data centre's total facility power divided by the power its IT equipment uses. A value near 1.0 means little overhead. It grades the building, not the work.
3 minute read - Guide 03
Tokens per watt: what it measures and how to read it
Tokens per watt is inference output divided by the power drawn to produce it. It shows what each unit of energy buys, at the point where the watts were read.
3 minute read - Guide 04
AI for energy in data centres
AI acts on data centre energy in two places. On the building side it tunes cooling and distribution. On the load side it adjusts the power the hardware draws.
4 minute read - Guide 05
How to reduce GPU power consumption
Match the draw to the work: pack work onto fewer GPUs, tune serving settings, and lower each GPU's clock when the work allows, checking throughput holds.
4 minute read - Guide 06
Data centre power capacity
Power capacity is the fixed envelope set by a grid connection or colocation contract. What matters is how much compute fits inside it, and that depends on where you count.
4 minute read
Then your own numbers
Every guide ends at the same question: what fits on your site.
The instrument answers it on your own hardware and the power you hold. The case studies work it through common situations.