Meganeura runs on any device with Vulkan or Metal support. Performance is best when the GPU and driver support hardware-accelerated cooperative matrix operations, but the library functions correctly without them — including on CPU via LavaPipe.
Rust version
Meganeura requires Rust 1.88 or later.
rustc --version # should print rustc 1.88.0 or higher
rustup update # update if needed
Vulkan
Meganeura uses Vulkan on Linux, Windows, and Android.
Cooperative matrix operations (VK_KHR_cooperative_matrix) enable hardware-accelerated 8×8 tile math for matmul-heavy workloads. The following GPUs support this extension:
| Vendor | Minimum hardware |
|---|
| NVIDIA | Volta architecture (GeForce RTX 20 series+, Quadro V100+) |
| AMD | RDNA3 architecture (Radeon RX 7000 series+) |
| Intel | Arc graphics (A-series+) |
Check whether your GPU supports VK_KHR_cooperative_matrix by running vulkaninfo | grep cooperativeMatrix or using the Vulkan Hardware Database. Runs on any Vulkan device
Meganeura works on any GPU with a Vulkan driver, even without cooperative matrix support. This includes older NVIDIA, AMD, and Intel GPUs that predate the hardware requirements listed above.
LavaPipe (software rendering)
LavaPipe is Mesa’s software Vulkan implementation. Meganeura runs on LavaPipe for development, testing, or environments without a GPU. Performance is substantially lower than hardware execution.
# Force LavaPipe on Linux
LIBGL_ALWAYS_SOFTWARE=1 cargo run --example mnist
LavaPipe is suitable for functional testing only. Do not use it for benchmarking or production inference.
- Apple M1 or later (all variants: M1, M1 Pro, M1 Max, M1 Ultra, M2, M3, M4, and their variants)
All Apple Silicon Macs ship with Metal support and simdgroup matrix instructions. No additional drivers are required.
macOS version
Apple Silicon Macs running macOS 12 Monterey or later are supported. Metal availability is tied to the OS version, not the hardware.
| Platform | Backend | Notes |
|---|
| Linux | Vulkan | NVIDIA, AMD, Intel, and LavaPipe |
| Windows | Vulkan | NVIDIA, AMD, Intel |
| macOS | Metal | Apple Silicon (M1+) recommended |
| iOS | Metal | Apple Silicon |
| Android | Vulkan | Device-dependent driver quality |
The benchmark table in the README shows measured forward and backward pass times for several GPUs. Meganeura outperforms PyTorch ROCm on AMD laptop APUs and matches PyTorch CUDA on mid-range NVIDIA hardware.
What is not supported
- CUDA — Meganeura does not use CUDA. Use the Vulkan backend on NVIDIA GPUs.
- ROCm — Meganeura does not use ROCm. Use the Vulkan backend with AMD RDNA drivers (RADV on Linux recommended).
- OpenCL — Not supported.
- x86/ARM CPU only — You can run on CPU via LavaPipe (Vulkan) but there is no native CPU backend.