CUDA Requirements
| Requirement | Detail | Verified |
|---|---|---|
| Operating system | Linux, x86-64 | Yes |
| GPU | NVIDIA, with a known compute capability | RTX 3050 Laptop GPU, compute capability 8.6; RTX 4060 Laptop GPU, compute capability 8.9 (built for 8.6) |
| Driver | Supporting your toolkit | 595.91.07 (advertises CUDA 13.2) with toolkit 13.4; 566.14 (advertises CUDA 12.7) with toolkit 12.4 |
| CUDA toolkit | 12.0 or newer: nvcc, CUDA runtime headers and library, all from the same toolkit | 13.4.92 and 12.4.131 |
| Host compiler | One your nvcc accepts | GCC 15.2.0 with nvcc 13.4; GCC 13.4 with nvcc 12.4 |
| CMake | 3.25 or newer | 4.2.3 |
| Build setting | CMAKE_CUDA_ARCHITECTURES set to your GPU | 86 |
GPUs
- Code built for compute capability 8.6 has been tested on 8.6 and 8.9 GPUs. No build for another architecture has been tested. The code uses no 8.6-specific feature, but an untested GPU is an untested GPU.
- Only one device is used.
- Older GPUs than the architecture you compile for cannot run the kernels; compile for your own GPU's capability.
Toolkit and compiler
CUDA 12.0 is the minimum because MENOS compiles its device code as C++20, which nvcc accepts only from 12.0 on. Configuring with an older nvcc stops with an error. CMake also refuses a build whose nvcc, CUDA headers and CUDA runtime library come from different major versions; that combination would otherwise build without complaint and misread device properties.
Your driver decides which toolkit you can run: a program built with a toolkit newer than the driver supports cannot use the GPU, and every CUDA test skips. For example, driver 566.14 runs CUDA 12.x builds but not CUDA 13 builds. Check nvidia-smi and pick a toolkit no newer than the CUDA version it reports, or one within the same major version (minor-version compatibility).
nvcc accepts only host compilers up to a certain version. CUDA 13.4's nvcc accepted GCC 15.2; the 12.4 toolkit accepts only GCC 13 or earlier. With an older toolkit, compile the whole build with a GCC it accepts (set both CMAKE_CXX_COMPILER and CMAKE_CUDA_HOST_COMPILER), as the cuda12-dev and cuda12-release presets do with GCC 13. GCC 12 is the oldest compiler MENOS supports. Pass the toolkit explicitly (CUDA Installation).
Warning flags
MENOS compiles its device code with the same warning discipline as the host code, with two flags dropped for CUDA sources because they fire inside nvcc's own generated files and CUDA's headers rather than MENOS code. This matters only if you develop MENOS.
Device memory
The device must have room for the state: 16 · 2ⁿ bytes for an n-qubit dense region, taken from the free memory the driver reports. See GPU Memory.
Numerical behaviour
Device kernels are compiled with -fmad=false so results match the CPU bitwise. This is part of the build; you do not configure it.
Platforms without CUDA support
CUDA is not supported on Linux aarch64 or macOS, and AMD (ROCm) and Apple (Metal) GPUs are not supported at all.