Results show that a compiler-defined symbolic interface turns native kernel synthesis into verified design-space exploration: compiler infrastructure preserves legality and correctness, while neural guidance improves workload-specific performance with lower search cost and complete coverage.
Abstract
Emerging accelerators often lack mature compiler backends, motivating neural agents that generate and repair kernels from architectural documentation and simulator feedback. This approach repeatedly reconstructs workload-invariant machine semantics--including instruction behavior, legality constraints, synchronization rules, and memory protocols--for every workload. We argue that these semantics should be compiled once into a persistent symbolic artifact, while neural reasoning should focus on workload-dependent mapping decisions. We present Zomboss, a compiler-mediated agentic framework for kernel generation that places neural search within a verified compiler boundary. Zomboss compiles machine semantics and legality constraints into a reusable mapping interface, then uses a neural agent to optimize workload-dependent decisions within the validated mapping space. Across 20 Gemmini and 36 PLENA workload instances, Zomboss returns a correct verified kernel on all 56 instances. Relative to the compiler default, Zomboss achieves geometric-mean speedups of $3.34\times$ on Gemmini and $1.10\times$ on PLENA. Relative to direct agentic generation, it reduces inference tokens by 71.2% on Gemmini and 54.2% on PLENA. These results show that a compiler-defined symbolic interface turns native kernel synthesis into verified design-space exploration: compiler infrastructure preserves legality and correctness, while neural guidance improves workload-specific performance with lower search cost and complete coverage.
CAKE, a compiler-agent co-design in which agents author CAKE IR, a typed, hardware-explicit schedule representation, exposes warp roles, memory movement, synchronization, and pipelines while supporting verification, cost modeling, and localized diagnostics.
LLM4LLM is introduced, a deployment-aware closed-loop optimization framework that starts from a target inference script, extracts phase-aware optimization tasks, searches with an experience-guided episodic agent, and accepts patches through in-model validation.
Hui Zeng, Pengfei Yang, Yanxin Chen et al.· 0 citations
Designing and authoring high-performance custom kernels for accelerators is a complex task that requires deep hardware-level expertise. Large Language Models (LLM) can be leveraged together with real-time compiler feedback to build agentic systems for kernel generation. In this work, we present MaxKernel, a multi-agent system that implements three distinct paradigms for TPU kernel development: (1) a Human-in-the-Loop (HITL) agent for collaborative, step-by-step design; (2) an Autonomous (Auto) agent that executes a fully automated, metric/trace-driven optimization loop; and (3) a Graph-Based Autonomous Search that scales the Auto agent for global exploration of the design space. All three paradigms leverage a shared pool of specialized sub-agents to handle planning, implementation, self-debugging, testing, and hardware profiling. We evaluate MaxKernel on JaxBench, a comprehensive suite of 50 diverse kernel tasks for TPUs, alongside complex, real-world workloads from state-of-the-art open-source models. We demonstrate that MaxKernel consistently generates highly optimized implementations, matching expert hand-tuned baselines and delivering significant performance across the benchmark. Our agent is open-sourced and available https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxKernel.
Shang-Kun Wang, Nina Cai, Charles Hoong et al.· 0 citations
A domain-specific debug agent is presented that addresses three core challenges in autonomous repair: mitigating knowledge scarcity through retrieved patterns and diagnostic instrumentation, ensuring integrity through anti-cheat detection and full-coverage evaluation, and controlling cost via convergence guards and bounded iteration.
Yansong Sun, Shenxi Wu, Siyuan Chen et al.· 0 citations
The Agent-Ass kernels outperform the Full-Agent artifacts across the evaluated definitions, indicating that expert-provided optimization directions, high-quality references, and workload context remain critical for reliable AI-driven kernel optimization.
Yue Shui, Chenyu Ma, Hang Xu et al.· arXiv.org· 1 citation
Modern AI systems depend on specialized accelerator kernels, whose development is complicated by increasingly diverse operators and hardware. LLMs and agentic systems promise to automate this work, but existing evaluations do not show whether their performance transfers across operator sources and hardware platforms, or what such transfer costs. We present KernelGenBench, the first unified multi-source and multi-chip infrastructure for evaluating LLM- and agent-generated Triton kernels. With a common Triton target spanning six hardware platforms, it provides the broadest cross-vendor hardware coverage among existing kernel-generation benchmarks. We report two controlled analytical views: KernelGenBench-MS (Multi-Source) covers 210 operators from PyTorch ATen, production vLLM operators, and proprietary cuBLAS routines, while KernelGenBench-MC (Multi-Chip) evaluates a semantically stable 110-operator subset across six hardware platforms. Our evaluation consumed over 15 billion tokens. Agentic execution improved correctness, but no method dominated across sources and platforms: vLLM posed the strongest correctness challenge, cuBLAS set the highest performance ceiling, and AutoKernel accuracy fell from 87% on NVIDIA to 25% on Iluvatar CoreX. These improvements were costly: specialized agents averaged 4.99 million tokens per successful operator, rising to 6.25 million for CUDA Optimized Skill. The results establish operator source, hardware platform, and agentic scaffold as distinct dimensions of kernel-generation capability, and show that success in a familiar source-hardware setting is not a reliable proxy for deployment readiness.
Pei-Yu Zang, Jian-Hang Tao, Jia-Ling Zhang et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
