Supported systems#

CI Tested systems#

We test the following systems in our CI servers:

  • Python 3.12-Python 3.14

  • Mac OS X 14 and 15

  • Ubuntu 22.04 and Ubuntu 24.04

  • Windows Server 2025

Supported systems:#

Operating systems#

  • Mac Silicon 14 or 15

  • Ubuntu 22.04 and 24.04 (x64 and ARM64)

  • Windows 10 or later (x86 only)

GPUs#

  • NVIDIA GPUs with compute capability 6.0 (sm_60, Pascal) through 12.0 (sm_120, Blackwell / Thor). Compute capability identifies the GPU features supported by an NVIDIA architecture; sm_60 and sm_120 are CUDA’s codes for these versions.

  • Metal GPUs

  • AMD GPUs

  • GPUs with Vulkan 1.2 or newer drivers (e.g. Intel Arc)

If you have a newer NVIDIA GPU (above sm_120), please open an issue on the Quadrants repo.

Backend / OS matrix#

Which backends are available on each supported platform. qd.cpu and qd.vulkan run on every OS; the other GPU backends are platform-specific because they wrap vendor drivers: CUDA on NVIDIA, ROCm (AMD’s GPU software platform) on AMD, and Metal on Apple.

OS \ backend

qd.cpu

qd.cuda

qd.amdgpu

qd.metal

qd.vulkan

macOS (Apple Silicon)

yes

n/a

n/a

yes

yes

Linux x64

yes

yes

yes

n/a

yes

Linux ARM64

yes

no

no

n/a

yes

Windows x86

yes

yes

no

n/a

yes

Windows ARM64

yes

no

no

n/a

yes

Notes:

  • qd.cuda requires an NVIDIA driver and CUDA runtime installed on the host. NVIDIA ships CUDA for Linux ARM64 and Windows ARM64, but quadrants does not support them yet.

  • qd.amdgpu requires HIP 6.0 or newer (HIP is AMD’s GPU programming interface) and currently wires up the Linux x64 ROCm path only. AMD’s GPU toolchain also ships on Windows and on some Linux ARM64 targets, but quadrants does not support them yet. On AMD GPUs, a subgroup is always 64 threads wide, so qd.simt.subgroup primitives operate over 64 lanes on both CDNA (Instinct) and RDNA (Radeon) hardware.

  • qd.metal is only available on Apple hardware and is the recommended GPU backend there.

  • qd.vulkan requires Vulkan 1.2 or newer. The Vulkan and Metal backends generate shaders in SPIR-V (an intermediate format for GPU programs), version 1.5 or newer; older targets are unsupported.

  • Quadrants normally detects GPU features and selects the shader format automatically. The qd.types.DeviceCapability class provides named constants for GPU features and shader-format versions when code specifies these choices explicitly. If your code uses its removed spirv_version_1_3 or spirv_version_1_4 constants, replace them with qd.types.DeviceCapability.spirv_version_1_5 to select SPIR-V 1.5.

  • Quadrants selects the Vulkan API version automatically. The vk_api_version argument to qd.init() has been removed; omit it from existing calls.

  • qd.vulkan on macOS bundles a copy of MoltenVK (a Vulkan-to-Metal translation layer) inside the wheel, so no separate install is required.

The Python backend (qd.python) has been removed. To run kernels on the CPU, use qd.init(arch=qd.cpu). CPU kernels are compiled; they do not execute as ordinary Python functions.