一个名为gpu-unlock的开源Linux内核模块被开发出来，可以解锁英伟达3000系列显卡的FP/BF16指令发布速率1。该模块支持GA102、GA103、GA104和GA106芯片的显卡，使其性能达到专业级显卡水平1。
该模块在NVIDIA驱动绑定前运行，将FP/BF16 FMA指令发布速度翻倍转换为FP32操作，无需在每次驱动更新时进行手动修补1。用户可以通过速度选择器进行配置，其中SS0设置为0x88888888可选择完整BF16速率，SS1设置为0x8 IMLA41。
该项目支持多个Linux发行版安装，包括Arch、Debian/Ubuntu、Fedora/RHEL等系统1。该模块还具备暂停-恢复功能，系统恢复时可自动重新解锁性能1。
An open-source Linux kernel module called gpu-unlock has been developed to unlock the FP/BF16 instruction issue rate on Nvidia's RTX 3000 series graphics cards 1. The module supports graphics cards based on the GA102, GA103, GA104, and GA106 chips 1. By doubling the FP/BF16 FMA instruction issue rate converted to FP32 operations, the tool enables consumer-grade GPUs to reach professional-level performance 1.
The gpu-unlock module operates by running before the Nvidia driver binds to the GPU, eliminating the need for manual driver patching with each driver update 1. Users can configure the instruction speed selector with settings such as SS0 0x88888888 to select full BF16 rate and SS1 0x8 IMLA4 1. The project supports installation across multiple Linux distributions including Arch, Debian/Ubuntu, and Fedora/RHEL 1. Additionally, the module provides suspend-resume functionality that automatically re-unlocks the GPU performance when the system resumes 1.
评论
还没有评论，欢迎留下第一条。