NVIDIA Launches CompileIQ in CUDA 13.3: AI-Powered Compiler Tuning Targets LLM Inference Performance

NVIDIA's new CompileIQ framework, launching in CUDA 13.3, uses AI-driven evolutionary algorithms to auto-tune compiler configurations for GPU workloads, targeting LLM inference hotspots like GEMMs and attention mechanisms that account for over 90% of compute, delivering measurable throughput gains already being deployed in production by leading AI labs.

• NVIDIA CompileIQ, now available in CUDA 13.3, is an AI-powered compiler auto-tuning framework that uses evolutionary and genetic algorithms to discover optimized internal compiler configurations tailored to specific GPU workloads, going beyond the default heuristics applied to all kernels.
• CompileIQ targets high-impact kernel hotspots such as GEMMs and attention mechanisms, which together account for over 90% of compute in LLM inference, where even fractional performance gains translate into significant overall throughput improvements.
• CompileIQ supports multi-objective optimization across runtime, compile time, and power consumption, producing portable and reproducible Advanced Controls Files that leading AI labs are already deploying in production for their most performance-critical workloads.