KernelSight-LM يضرب السقف 7 مرات في GPU Kernel Latency Prediction

KernelSight-LM predicts per-kernel GPU latency on unseen hardware to 3.8% error with just one calibration sweep — a 7.3x improvement over a comparable roofline baseline's 27.7% error.

How KernelSight-LM Decomposes LLM Inference

LLM inference couples serving-layer policies (prefix caching, continuous batching) with low-level GPU kernel execution. KernelSight-LM decomposes each serving step into four components: a roofline kernel model with a learned efficiency term, a communication model, a host-overhead model, and a discrete-event scheduler that captures caching and batching mechanics. That scheduler is what lets the simulator reproduce real-world serving behavior instead of treating each token as an independent event.

Two Prediction Tiers Trade Data for Accuracy

Two tiers let users choose based on available target-GPU data. The cross-generation tier uses no target-GPU measurements — just hardware specs and kernel microbenchmarks from previously profiled GPUs — and achieves 12.1% per-kernel error, a 1.8x improvement over the 22.0% roofline baseline. The target-measured tier adds one model-agnostic kernel-microbenchmark sweep on the target GPU, dropping per-kernel error to 3.8%.

End-to-End Errors Match Dedicated Profiling Tools

Across six model families, the cross-generation tier yields median errors of 15.4% for TTFT, 12.8% for TPOT, and 3.0% for throughput. The target-measured tier improves those to 14.3%, 6.2%, and 2.7% respectively. These numbers meet the accuracy of dedicated profiling tools while collecting far less on-device data.

KernelSight-LM’s kernel-level bottleneck breakdowns let engineers plan capacity and run hardware-software co-design experiments without deploying every model-variant on every GPU generation.

Source: KernelSight-LM: A Kernel-Level LLM Inference Simulator
Domain: arxiv.org