ELF: the Energy-Latency Frontier dataset for GPU inference

Abstract

A dense power-response dataset for AI inference on data-center GPUs. For 20 inference models (LLM decode, Stable-Diffusion UNet, ViT, and CNNs) across four NVIDIA GPUs (T4, L4, A10G, A100), it sweeps the board power limit from the driver floor to TDP, and, via a graphics-clock (DVFS) sweep, reaches below that floor, recording measured board power, achieved throughput, and a four-phase latency split, at a fixed batch and an auto-calibrated saturating batch, plus a high-rate power-cap actuation trace. The headline artifact is data/elf_master.csv: a consolidated master table with one row per measured operating point (card x workload x actuator setting), spanning both actuators (power cap and clock/DVFS) and both load regimes (saturated and non-saturated), fully labelled with provenance, saturation flags, canonical-subset flags, card specs, workload features, and fitted response-law parameters; every column is documented in data/elf_master_dictionary.md. The sweep collections hold 6,488 measured windows (two to three consecutive two-second windows per operating point); the first window after each cap change still carries the settling pause in NVML's trailing power average, so the analyses discard it (code/settle_filter.py), and the settled windows are consolidated into the 2636-row master table. The files keep every window. Companion to the paper 'The Joule Point: an Energy-Optimal Operating Point for AI Inference' (Apartsin and Aperstein). Includes the exact AWS measurement scripts, the consolidation and verification code, and the analysis and simulation code that computes every number and figure in the paper (code licensed MIT; data CC-BY-4.0). See README.md for schema and known limitations.