Get Started
Home
Topics
Search
Library
Research questionHow can LLM-generated GPU kernels remain performant across hardware, inputs, and programming models?LLM-generated kernels may not exploit architecture-specific parallelism, and optimizations that work for one device or input may not transfer to others. Standard prompting and profiling feedback alone leave a large search space for finding effective implementations.
AI
Code Generation & Program Synthesis
Evaluation & Benchmarks
Neural and Evolutionary Computing
Research Paper
Technology
Latest papersRecent research connected to this question, newest first.KernelFoundry: Hardware-aware evolutionary GPU kernel optimizationThe evidence concerns SYCL kernel generation, with CUDA kernels used for comparison, evaluated on Kernel-Bench, robust-kbench, and custom tasks. The described system supports remote benchmarking across diverse hardware and reports an average 2.3× SYCL speedup on Kernel-Bench; the input does not establish broader deployment results.research paper · Sep 3, 2026
Related questions
How can TPU kernels be generated and optimized without scarce hardware expertise?How can LLMs reliably construct analytical hardware-performance models from workload, architecture, and mapping specifications?How can GPUs efficiently compute unnormalized Gaussian kernel sums with signed weights without custom kernels?How can symbolic regression optimize constants efficiently across heterogeneous expression-tree populations on GPUs?