Skip to content

gpu.md

Latest commit

 

History

History
111 lines (81 loc) · 2.05 KB

gpu.md

File metadata and controls

111 lines (81 loc) · 2.05 KB

GPU Virtualization in AetherVM

Overview

AetherVM provides two GPU virtualization approaches:

  1. Virtual GPU (vGPU): Software-emulated GPU using WebGPU/Vulkan
  2. GPU Passthrough: Direct hardware access via VFIO

Virtual GPU

Architecture

The virtual GPU uses WebGPU as its backend, providing:

  • Cross-platform compatibility
  • Shader compilation and execution
  • Graphics pipeline emulation
  • Compute shader support

Usage

use aethervm_gpu::vgpu::VirtualGpu;

let mut vgpu = VirtualGpu::new("default-gpu");
vgpu.init().await?;

GPU Passthrough

Requirements

  • IOMMU enabled in BIOS
  • VT-d (Intel) or AMD-Vi support
  • GPU in separate IOMMU group
  • Host driver unbound from device

Configuration

[gpu]
enabled = true
passthrough = true
device = "0000:01:00.0"
vendor_id = "10de"  # NVIDIA

Setup on Linux

# Enable IOMMU in kernel parameters
# Add to /etc/default/grub:
GRUB_CMDLINE_LINUX="intel_iommu=on iommu=pt"

# Update grub
sudo update-grub

# Unbind GPU from host driver
echo "0000:01:00.0" > /sys/bus/pci/drivers/nvidia/unbind

# Bind to vfio-pci
echo "10de 1b80" > /sys/bus/pci/drivers/vfio-pci/new_id

Supported Workloads

  • Graphics: OpenGL, Vulkan, DirectX
  • Compute: CUDA, OpenCL, ROCm
  • AI/ML: TensorFlow, PyTorch GPU acceleration
  • Gaming: Cloud gaming, game streaming

Performance

Mode Latency Throughput Compatibility
vGPU Medium Medium High
Passthrough Low Native Medium

Examples

CUDA Workload

let vm = AgentVM::builder()
    .name("cuda-vm")
    .cpu_cores(8)
    .memory_gb(16)
    .enable_gpu(true)
    .build()
    .await?;

// GPU is available inside VM
vm.execute_agent_script(r#"
    // Run CUDA code
    cuda_available()
"#).await?;

Gaming VM

let vm = AgentVM::builder()
    .name("gaming-vm")
    .cpu_cores(6)
    .memory_gb(16)
    .enable_gpu(true)
    .build()
    .await?;

// Optimized for low latency