Chip for AI14 resources
Kernel Optimization
Kernel generation, optimization, evaluation, and compiler tools for GPUs, CPUs, and accelerators.
Generation, Optimization, and Evaluation
- AccelOpt (2025/2026) — Builds optimization memory from slow-fast kernel pairs to autonomously optimize kernels on AWS Trainium, and provides NKIBench. Paper
- AKG kernel Agent (2025) — Uses multiple agents to generate, port, and tune kernels across GPUs and NPUs, supporting several kernel DSLs.
- Atrex Kernel Agent (2026) — Uses a GPU knowledge base, profiling tools, and structured workflows for kernel implementation, diagnosis, and iterative optimization.
- AWS Neuron Agentic Development — AWS’s official NKI agent tools cover kernel authoring, compilation repair, profiling, performance analysis, and API migration.
- CUDA Agent (2026) — Combines large-scale agentic RL, CUDA task data, expert skills, and verifiable execution environments to train high-performance kernel generation. Paper
- CUDA-L1 (2025/2026) — Optimizes CUDA kernels with contrastive reinforcement learning and execution feedback, discussing performance rewards and reward hacking. Paper
- KernelBench (2025) — Uses PyTorch tasks to evaluate whether models can generate correct and faster GPU kernels, with execution and measurement tools.
- Kevin-32B (2025) — Trains CUDA kernel generation and optimization through multi-turn reinforcement learning, enabling iterative code improvement with execution feedback. Paper Model
- MaxKernel (2026) — Provides human-agent collaboration, autonomous optimization, and graph-search modes for TPU kernels, combining compilation, testing, and hardware profiling. Code
- PyTorch KernelAgent / KernelFalcon (2025–2026) — Organizes kernel generation, numerical validation, hardware profiling, and bottleneck analysis into a multi-agent loop, with shared memory guiding optimization. Technical Blog
Kernel Libraries and Compilers
- Arm KleidiAI — An AI micro-kernel library for Arm CPUs, with architecture-optimized matrix operations and data-packing routines. Main Repository
- DeepEP — Provides high-throughput and low-latency communication paths for MoE expert parallelism, connecting model routing, GPU kernels, and network transport.
- TileLang — Expresses GPU, CPU, and accelerator kernels through tiles to simplify low-level optimization and implementation across hardware backends.
- Triton — A language and compiler for high-performance GPU kernels, used as a code-generation target by multiple kernel agents.