SOLAR Automates Speed-of-Light Performance Analysis for Deep Learning

Optimizing deep learning models for specific hardware often requires understanding their theoretical performance limits, known as Speed-of-Light (SOL) bounds. Traditionally, deriving these bounds is a manual, error-prone, and time-consuming process, making it difficult to keep pace with rapid model development. To address this, researchers have introduced SOLAR, an AI-powered framework that automates SOL analysis. SOLAR takes PyTorch and JAX source code as input and leverages both generative and deterministic components. An LLM frontend translates the source programs into an executable Affine Loop Intermediate Representation (IR), which is then validated by comparing outputs. A deterministic flow lifts this IR into an einsum graph, and an analytical backend subsequently computes various SOL bounds, including unfused, fused, and cache-aware limits. The framework offers comprehensive operator and language coverage, producing validated bounds with no observed SOL violations. It provides multi-fidelity analysis, allowing for tighter bounds and surfacing critical optimization insights. Evaluations across KernelBench, JAX/Flax models, and robotics workloads demonstrated SOLAR's utility in headroom analysis, identifying optimization opportunities, cross-platform exploration, and inverse-roofline hardware provisioning.