virtio-nvgpu是一个开源项目，通过在驱动ABI级别转发NVIDIA内核驱动的ioctl调用，使KVM虚拟机能够获得NVIDIA GPU的近原生访问能力1。该方案使虚拟机的渲染性能相比裸机仅相差2%以内，CPU开销与主机保持相同1。
该技术支持Vulkan、OpenGL、CUDA和NVENC编码1，适用于无头流媒体应用场景。性能测试数据显示，在813,691帧的测试中仅有13,792条消息跨越虚拟机边界，平均每59帧发生一次1。四个虚拟机共享一块RTX 3060时的总体性能为103.7 fps，而单个虚拟机的性能为102.9 fps1。项目支持NVIDIA驱动版本535.129.03、580.178.04、595.71.05及更新版本1，基准测试采用RTX 3060和驱动版本595.99.02进行1。
An open-source project called virtio-nvgpu enables near-native access to Nvidia GPUs from within KVM virtual machines by forwarding Nvidia kernel driver ioctl calls at the driver ABI level.1 The approach achieves rendering performance within 2 percent of bare metal while maintaining identical CPU overhead.1 The implementation supports Vulkan, OpenGL, CUDA, and NVENC encoding, making it suitable for headless streaming applications.1
Performance analysis demonstrates the efficiency of the virtualization layer.1 Across 813,691 frames, only 13,792 messages crossed the boundary between host and guest, averaging one crossing per 59 frames.1 When four virtual machines share a single RTX 3060 GPU, the combined throughput reaches 103.7 frames per second compared to 102.9 fps for a single virtual machine.1 The technology supports Nvidia driver versions 535.129.03, 580.178.04, 595.71.05, and later releases, with benchmarking conducted using driver version 595.99.02 on an RTX 3060.1
评论
还没有评论，欢迎留下第一条。