Filtered by: Hardware × Clear all

Are LLM-Generated GPU Kernels Production-Ready? A Trace-Driven Benchmark and Optimization Agent

Lingyun Yang, Yuxiao Wang, Shenghao Liang, Linfeng Yang 2026-07-19
Inference LLM GPU Hardware × Agent

Atrex-Bench addresses the problem that existing GPU kernel benchmarks use synthetic or curated workloads, not production traces. The method samples 30 operators and 440 shapes from full-cluster inference traces, weighting each by GPU time and card-hours. Experimental evidence shows the best vanilla model reaches only ~10% of roofline performance, with much apparent correctness from PyTorch fallbacks. This matters because the co-released Atrex-Kernel-Agent (AKA) converts fallbacks into kernels matching hand-tuned baselines, demonstrating a path to production-ready LLM-generated kernels.

PDF

Valinor: Architectural Support for Fast, Energy-Efficient and Programmable Physical Memory Allocation

Konstantinos Kanellopoulos, Spiros Galanopoulos, Konstantinos Sgouras, Vlad-Petru Nitu 2026-07-19
Workload Runtime Hardware × Simulation

The problem is that physical memory allocation in current systems incurs high overhead from minor page faults, which can account for up to 54% of runtime and 40% of system energy in short-lived workloads like serverless functions. The method, Valinor, is a hardware-OS cooperative substrate that introduces a programmable hardware allocation engine executing compact OS-supplied allocation libraries at near fixed-hardware speed. On a BOOM RISC-V soft core running Linux, Valinor accelerates allocation by 17x, improves end-to-end performance by 16%, and reduces energy consumption by up to 8%, with full-system simulation confirming hardware-class performance across six allocation libraries. This matters because Valinor provides the flexibility to support diverse allocation policies and adapt to new hardware conditions while achieving the performance and energy efficiency of dedicated hardware.

PDF