Compute performance per watt compares output with power use under defined conditions. For AI inference, the output might be tokens per second, completed requests, or another workload-specific measure.
Meaningful comparisons use the same model, quality target, latency, batch size, precision, and power boundary. A headline ratio can mislead when one test omits host, memory, cooling, or idle consumption.
ELI5
Compute performance per watt measures how much useful computing work a system completes for the electricity it uses. A higher value means the system gets more work from the same amount of power under comparable conditions.
For example, two servers might answer the same number of AI requests each second, but one uses half the electricity. That server has better performance per watt, provided both tests use the same model, answer quality, and response-time target.
Why does performance per watt matter for AI?
Energy affects operating cost, cooling, data-center capacity, deployment density, and the environmental impact of serving models at scale.
Can performance-per-watt results be compared directly?
Only when the workload, output quality, latency, measurement boundary, hardware configuration, and software stack are sufficiently comparable.
