A leading processor manufacturer has published a detailed breakdown of the energy its newest data-centre chip consumes per computing task, a level of disclosure that remains rare across the industry.
Why the detail matters
Operators of large computing facilities are under growing pressure to account for electricity use as demand for AI workloads rises.
“Per-task energy is the number that actually matters to an operator's bill. It is good to see it published, even if the methods still differ.”
Several industry groups are working toward a common reporting standard.
▦ Reported energy per standard task
| Workload | Energy | vs. prior gen |
|---|---|---|
| Inference (text) | 0.9 Wh | −28% |
| Training step | 4.2 Wh | −19% |
| Idle (per hour) | 11 W | −12% |
This article is free to read. It always will be — no paywall, no account, no tracking.