Xe-Forge: Multi-Stage LLM-Powered Kernel Optimization for Intel GPU

Xe-Forge is an LLM-powered system that automates kernel optimization for Intel GPUs, eliminating repetitive manual porting work that typically gates algorithm deployment on new accelerators. Testing on 97 kernels achieved 1.17x geometric mean speedup with 67% of kernels improving and some exceeding 5x gains, demonstrating that structured domain knowledge combined with hardware-in-the-loop verification can systematically accelerate hardware adoption.

Xe-Forge addresses a critical friction point in hardware acceleration: the labor-intensive process of porting optimized kernels across different GPU architectures. Developers currently face a repetitive cycle of profiling, tuning, and architecture-specific modifications for each new hardware platform, significantly delaying time-to-market for algorithms on emerging accelerators like Intel Arc GPUs. This system automates nine optimization stages using a Chain-of-Verification-and-Refinement agent that generates candidates, validates them on actual hardware, and iterates on failures—eliminating guesswork through real-world feedback.

The technical approach combines LLM reasoning with domain-specific constraints encoded in a curated knowledge base capturing Intel GPU particularities absent from standard training data. This hybrid methodology prevents the model from proposing architecturally invalid optimizations while maintaining flexibility for novel improvements. The 1.17x geometric mean speedup across 97 Level-2 kernels may seem modest, but achieving 67% improvement rate with nine kernels exceeding 5x gains and Flash Attention reaching 82x speedup demonstrates significant heterogeneous outcomes rather than uniform gains.

For the broader AI hardware ecosystem, this work signals that systematic automation of kernel optimization is viable and essential for competing accelerators to capture developer mindshare. Intel's Arc GPUs face adoption barriers against entrenched NVIDIA dominance partly due to optimization friction; reducing that friction algorithmically could accelerate ecosystem growth. The open-ended discovery stage suggests the system can identify non-obvious optimizations beyond known patterns, potentially uncovering hardware capabilities developers miss. Future implementations may generalize this approach to other architectures, fundamentally changing how hardware vendors support algorithm deployment.

• →Xe-Forge automates kernel optimization for Intel GPUs using LLM agents with hardware-in-the-loop verification, reducing manual porting effort.
• →Achieved 1.17x geometric mean speedup with 67% kernel improvement rate and individual kernels reaching up to 82x speedup.
• →Incorporates domain-specific constraints in a curated knowledge base to keep optimizations within architectural validity bounds.
• →Demonstrates that structured domain knowledge combined with real hardware validation can systematically eliminate deployment friction for new accelerators.
• →Potential to accelerate Intel GPU adoption by reducing the optimization barrier that currently favors established NVIDIA platforms.