Question 172·200 Super-Hard SAT Reading Questions·Information and Ideas
Engineers often estimate the battery cost of running an artificial intelligence model by counting the computational operations it performs. Computer engineer Priya Raman and her colleagues tested this approach using two speech-recognition models on smartphones. A compressed model performed 35 percent fewer operations than a standard model but used only 10 percent less energy in scripted trials. Diagnostic logs indicated that the compressed model repeatedly transferred small pieces of data from memory, activating energy-intensive hardware. Without changing the model's calculations, the team rearranged its stored data so successive steps could reuse pieces already loaded; energy use then fell by an additional 22 percent, while accuracy remained constant. In later reports, the researchers retained operation counts but presented them alongside measurements of data transferred between memory and the processor.
Which choice best states the main idea of the text?
For a main-idea question, trace the passage's full progression: the initial assumption, the result that complicates it, the explanation for that result, and any final qualification. Choose the option that incorporates this complete relationship without exaggerating a limited finding or contradicting a stated detail.
Hints
Compare prediction with outcome
Consider why a 35 percent reduction in operations produced a much smaller initial reduction in energy use.
Focus on the intervention
Notice what the researchers changed before the additional energy reduction occurred—and what they did not change.
Include the final qualification
The researchers retained operation counts rather than rejecting them. Look for an answer that reflects both their usefulness and their limitation.
Step-by-step Explanation
Identify the initial expectation
Because the compressed model performed 35 percent fewer operations, operation count suggested that it should use substantially less energy.
Examine the observed results and explanation
The model used only 10 percent less energy at first because repeatedly transferring data activated energy-intensive memory hardware. After the data were rearranged to reduce those transfers, energy use fell further without any change in accuracy or calculations.
Account for the qualification
The researchers continued to report operation counts, so the text does not dismiss that measure. Instead, their decision to report memory traffic as well indicates that computational work alone does not fully determine battery savings.