mvgal-docs

Documentation for MVGAL — Multi-Vendor GPU Aggregation Layer for Linux.

What is MVGAL?

Most Linux systems with multiple GPUs (e.g. an AMD RX 7900 + NVIDIA RTX 4080) treat each card as a completely separate device. Applications can only use one at a time, leaving the other idle.

MVGAL solves this by aggregating all available GPUs — regardless of vendor — into a single logical device. Any application, game, or compute workload can use it without modification.

flowchart TD
    A["Your Application / Game / AI Workload"]

    A --> B["Vulkan"]
    A --> C["OpenCL"]
    A --> D["CUDA"]

    B --> E
    C --> E
    D --> E

    subgraph Interception["MVGAL API Interception Layer"]
        E["VK_LAYER_MVGAL<br/>libmvgal_opencl<br/>libmvgal_cuda"]
    end

    E -->|Unix socket| F

    subgraph Daemon["mvgald (daemon)"]
        F["Scheduler<br/>MemoryMgr<br/>PowerMgr<br/>Metrics<br/>IPC"]
    end

    F --> G["amdgpu.ko<br/>(AMD GPU)"]
    F --> H["nvidia.ko<br/>(NVIDIA GPU)"]
    F --> I["i915 / xe.ko<br/>(Intel GPU)"]
    F --> J["mtgpu-drv.ko<br/>(MTT GPU)"]

Loading

Features

• Heterogeneous multi-GPU — AMD, NVIDIA, Intel, and Moore Threads GPUs in any combination
• Transparent interception — Vulkan layer, OpenCL ICD, CUDA shim; no application changes needed
• 7 scheduling strategies — Round-robin, AFR, SFR, Task-based, Compute offload, Hybrid, Single-GPU
• Unified memory manager — DMA-BUF zero-copy, PCIe P2P, host-RAM staging fallback
• GPU health monitoring — Temperature, utilization, VRAM pressure with configurable thresholds
• Steam/Proton integration — Frame pacing, AFR for games, DXVK and VKD3D-Proton compatible
• Power management — Idle detection, GPU parking, dynamic frequency scaling
• Memory-safe subsystems — Fence manager, memory tracker, capability model written in Rust
• Qt dashboard + REST API — Real-time monitoring, scheduler control, log viewer

Supported Hardware

Vendor	Architectures	Driver
AMD	RDNA 1/2/3, GCN, APU (Vega/RDNA)	amdgpu
NVIDIA	Turing (RTX 20xx), Ampere (RTX 30xx), Ada (RTX 40xx), Pascal	nvidia-open / proprietary
Intel	Gen 9–12 (iGPU), Xe / Arc (discrete)	i915 / xe
Moore Threads	MTT S60, S80, S2000	mtgpu-drv

Install from COPR (no build needed)

MVGAL is available as a pre-built package via Fedora COPR. No need to compile from source.

sudo dnf copr enable axogm/mvgal
sudo dnf install mvgal

Supported targets: Fedora 40+ · RHEL/AlmaLinux/Rocky 9 & 10 · CentOS Stream 9 & 10 · Amazon Linux 2023

Quick Start

# Start the daemon
pkexec systemctl start mvgald
pkexec systemctl enable mvgald   # start on boot

# Verify
mvgal-info          # list detected GPUs
mvgal-status        # real-time utilization
mvgal-compat --system   # check readiness

CLI Tools

Tool	Description
mvgal-info	Print all detected GPUs, VRAM, temperature, utilization
mvgal-status	Real-time GPU utilization/VRAM bars; --watch for continuous refresh
mvgal-bench	Memory bandwidth, compute FLOPS, scheduling latency
mvgal-compat	System readiness check + per-app compatibility database
mvgal-config	Configure scheduler mode, idle thresholds, GPU enable/disable
mvgal	Main CLI: start/stop daemon, set strategy, show stats

Scheduling Strategies

Strategy	Best For
Round-robin	Even distribution, general compute
Alternate Frame Rendering (AFR)	Gaming — odd/even frames on different GPUs
Split Frame Rendering (SFR)	Gaming — horizontal/vertical tile split
Task-based	Mixed graphics + compute workloads
Compute offload	AI/HPC — route compute to best GPU
Hybrid adaptive	Automatic selection based on workload metrics
Single GPU	Fallback / compatibility mode

Steam / Proton Integration

Add to Steam launch options:

ENABLE_MVGAL=1 MVGAL_STRATEGY=afr %command%

Variable	Values	Description
ENABLE_MVGAL	0 / 1	Enable MVGAL for this launch
MVGAL_STRATEGY	afr, sfr, hybrid, single	Scheduling strategy
MVGAL_FRAME_PACING	0 / 1	Enable vsync-aligned frame pacing
MVGAL_GPU_MASK	hex bitmask	Which GPUs to use (e.g. 0x3 = GPU 0+1)

Memory Management

MVGAL uses a three-tier transfer strategy:

1. DMA-BUF zero-copy (preferred — kernel-supported, all vendors)
2. PCIe P2P transfer (fallback — requires same root complex)
3. Host-RAM staging (last resort — always works, highest latency)

License

• Kernel module: GPL-2.0-only
• Userspace components: GPL-3.0-only
• Rust crates: MIT OR Apache-2.0

Source Code

The MVGAL source code is available for purchase at:

• Patreon Shop — patreon.com/axogm/shop
• Ko-fi Shop — ko-fi.com/axogm/shop

Your purchase helps fund continued development of MVGAL and other open-source projects.

Getting Help

• Issues & bugs — github.com/TheCreateGM/mvgal-docs/issues
• Suggestions & improvements — github.com/TheCreateGM/mvgal-docs/discussions