An AI agent named ForgeKernel has autonomously generated optimized code for AI hardware, surpassing human-written benchmarks. In a 22.5-hour live demonstration, ForgeKernel wrote code that was 9.8% faster than DeepSeek's FlashMLA Prefill operator on an Nvidia B200 GPU. This AI-generated operator achieved superior performance across multiple benchmarks on Ascend hardware, demonstrating its adaptability to different architectures. The system, developed by Face-to-Face Intelligence and OpenBMB, represents a significant step towards AI-driven code generation and optimization, potentially accelerating the development and efficiency of AI infrastructure, particularly for domestic Chinese hardware. AI
IMPACT Accelerates AI infrastructure development and efficiency, particularly for domestic hardware, by automating complex operator optimization.
RANK_REASON AI agent autonomously generates optimized code that surpasses human-written benchmarks, representing a novel capability in AI-driven development. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- arXiv
- Ascend 910C
- Ascend 950DT
- Ascend 950PR
- CANNBench
- CUDA
- DeepInfer: Open-Source Deep Learning Deployment Toolkit for Image-Guided Therapy
- DeepSeek
- FlashMLA
- ForgeKernel
- JustGoal
- MIT
- Nvidia B200
- OpenBMB
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
