8 serial calls burn 4.86x GPU energy vs one batched call in test-time scaling

iowastate · hf · 2026-09-19

This paper shows candidate count N alone doesn't describe the systems cost of test-time scaling. On 500 GSM8K prompts, raising N from 1 to 8 lifted accuracy by 8.4pp (Phi-3-mini) and 18.4pp (Qwen2.5-1.5B). But fixing N=8 and comparing schedules (1x8, 2x4, 4x2, 8x1), eight serial A100 calls used 4.64–4.86x the gross GPU energy and had 5.77–6.12x the P95 latency of one batched call — consistent across three nodes and a SciQ/V100 setup. When candidates are independent and memory allows, fewer calls with larger batches are far more efficient; evaluations should report schedule and GPU-level metrics alongside accuracy.

Original post →

More from Infra

Infra channel →