Measure energy per completed workload.
Energy research for data centers, HPC labs, and cloud operators
An efficiency evaluation measures energy per correctly completed workload alongside throughput and response time. The published observation below concerns package energy in one test; its original run conditions are needed to interpret it.
One low-load point improved two measures.
One published low-load test summary reports 1.154x throughput and 0.875x package energy per correct work.
The same point worsened other measures.
P95 latency worsened and adjusted package power increased. Other load profiles were mixed or inconclusive.
Package energy isn’t the data-center bill.
The current evidence doesn’t establish total wall energy, cooling cost, thermal throttling, or lower cost per token.
Broader energy claims require new matched runs.
The next campaign must name the workload, useful work, energy boundary, thermal state, and confidence interval.
Working through a hardware or runtime decision? Explore compute evaluations
Startup and discovery, file and project organization, Forge libraries, Taskman, agent work and Kestowv pilot operations.