Accelerator
GPU, TPU, NPU or other inference accelerator energy.
MEASUREMENT FRAMEWORK / V0.1
Useful Intelligence
per Joule.
A proposed framework for measuring how much useful AI performance is obtained from the physical energy consumed to produce it.
UI/J
Useful Intelligence
Measured Energy
01 / MOTIVATION
Performance alone is not enough.
AI systems are commonly evaluated using accuracy, task scores, latency, throughput and economic cost. These measurements are essential, but they do not directly describe how efficiently a system converts physical energy into useful computational results.
LowJoule proposes treating energy as an explicit experimental variable alongside model quality.
Efficiency = Useful Output / Physical Energy
02 / GENERAL DEFINITION
A quality-normalized energy metric.
Let Q denote a task-specific quality function and E the measured energy required to execute the evaluated workload.
UI/J = E[Q] / E[E]
The numerator is deliberately task-dependent. Useful intelligence cannot be reduced to a single universal score without specifying what constitutes useful performance for the workload being studied.
Important distinction
UI/J is currently a proposed research framework, not an established SI unit or standardized industry benchmark. Comparisons are meaningful only when task definition, quality metric, workload and energy-measurement boundaries are controlled.
03 / RELATIVE UI/J
Compare systems under controlled conditions.
For many experiments, relative efficiency is more informative than an isolated absolute score. Let π₀ represent a baseline policy and π₁ an experimental policy.
RUI/J = (UI/J)π₁ / (UI/J)π₀
A value greater than one indicates improved quality-normalized energy efficiency relative to the selected baseline under the same experimental protocol.
04 / QUALITY-CONSTRAINED SAVINGS
Energy reduction without hiding quality loss.
A system can trivially consume less energy by performing less useful work. Energy savings should therefore be evaluated subject to an explicit quality constraint.
SE = 1 − Eadaptive / Ebaseline
subject to
Qadaptive ≥ Qbaseline − ε
Qadaptive ≥ Qbaseline − ε
The tolerance ε must be declared before interpreting an energy reduction as an efficiency improvement.
05 / MEASUREMENT BOUNDARY
What energy are we measuring?
Energy measurements depend strongly on system boundaries. LowJoule experiments should therefore explicitly declare which components are included.
CPU
Host-side computation and orchestration.
Memory
Memory access and data movement where measurable.
System
Whole-node power when physical instrumentation is available.
Communication
Network or interconnect costs when relevant to the experiment.
Infrastructure
Facility overhead may be reported separately when reliable data is available.
06 / EXPERIMENTAL PROTOCOL
A reproducible UI/J experiment.
01
Define workload
Fix the dataset, prompts, task distribution and evaluation procedure.
02
Define quality
Select an explicit metric Q appropriate to the task.
03
Fix hardware
Record accelerator, CPU, memory, software stack and runtime configuration.
04
Measure baseline
Execute the fixed-compute policy and record quality, latency and energy.
05
Measure adaptive policy
Repeat the same workload using the LowJoule policy.
06
Report uncertainty
Repeat measurements and publish variance, confidence intervals and measurement assumptions.
07 / MINIMUM REPORTING STANDARD
Every result should be auditable.
ModelModel + exact version
HardwareAccelerator / CPU / RAM
WorkloadDataset + sample count
QualityMetric Q + evaluation method
EnergyJoules + measurement method
LatencyMean + distribution
PolicyFixed or adaptive
UncertaintyRepeated runs + variance
08 / BENCHMARK VISION
From metric to open benchmark.
The long-term objective is an open LowJoule benchmark in which models and computational policies can be compared across explicit quality-energy Pareto frontiers.
P = {(Eπ, Qπ) : π ∈ Π}
Rather than declaring a single universally optimal system, the benchmark would expose the trade-off between quality and physical energy across policies, models and hardware configurations.
09 / TARGET
The objective is not simply to make AI consume less energy. The objective is to eliminate computation whose expected utility does not justify its physical cost.