Build fails on NVIDIA Blackwell (sm_120, CC 12.0) with CUDA 13.2 — patches required

[dmeliksetian]

Environment

• GPU: NVIDIA RTX PRO 4500 Blackwell (Compute Capability 12.0)
• Driver: 595.45.04
• CUDA: 13.2
• OS: Ubuntu (GCC 13.3.0)
• Qiskit Aer: 0.17.2
• Python: 3.12.3

Problem

Building qiskit-aer from source fails with four separate issues when targeting
NVIDIA Blackwell (sm_120 / CC 12.0) with CUDA 13.2. All are caused by the source
predating the Blackwell architecture and CUDA 13.x.

Failures and fixes

1. select_compute_arch.cmake missing Blackwell

cuda_select_nvcc_arch_flags has a hardcoded arch list that stops at Ampere (8.6).
Fix: add version blocks for Hopper (11.8), Ada (12.0), and Blackwell (12.8) to the
known arch list and the named-arch elseif chain.

2. CMakeLists.txt — cuda_select_nvcc_arch_flags doesn't accept 12.0

Even after patching the cmake file, the dotted version regex only matches single
digits (e.g. 8.6), so 12.0 falls through to the unknown-arch error.
Fix: bypass cuda_select_nvcc_arch_flags entirely and hardcode:

set(AER_CUDA_ARCH_FLAGS "-gencode arch=compute_120,code=sm_120")
set(AER_CUDA_ARCHITECTURES "120")
set(CMAKE_CUDA_ARCHITECTURES "120")

3. C++14 incompatible with CUDA 13.2 Thrust/CCCL

CUDA 13.2 bundles Thrust/CCCL which requires C++17. Aer defaults to C++14.
Fix: pass -DCMAKE_CXX_STANDARD=17 -DCMAKE_CUDA_STANDARD=17 at build time,
or update CMakeLists.txt to set these when CUDA >= 12.

4. thrust::unary_function removed in Thrust 2.x

src/simulators/statevector/chunk/thrust_kernels.hpp inherits from
thrust::unary_function and thrust::binary_function, which were removed
in Thrust 2.x (bundled with CUDA 12+).
Fix: add a compatibility shim after #include "misc/wrap_thrust.hpp":

namespace thrust {
  template<typename Arg, typename Result>
  struct unary_function {
    typedef Arg argument_type;
    typedef Result result_type;
  };
  template<typename Arg1, typename Arg2, typename Result>
  struct binary_function {
    typedef Arg1 first_argument_type;
    typedef Arg2 second_argument_type;
    typedef Result result_type;
  };
}

Suggested fix

• Update the minimum C++ standard to 17 when CUDA >= 12 is detected
• Add Blackwell/Hopper/Ada to select_compute_arch.cmake
• Replace thrust::unary_function with a bundled shim or std:: equivalent
• Consider replacing the legacy FindCUDA module with modern find_package(CUDAToolkit)

[dmeliksetian]

Hi team,

I'm a Qiskit Advocate and I've already worked through all four of these fixes on my end to get the build working. Happy to put together a proper PR with the patches if that would be helpful — just let me know and I'll get it submitted.

[spital]

Hi team, another Qiskit Advocate here, upgraded qiskit-aer-gpu for Conan 2 and Blackwell / sm120 and Cuda 13 support few weeks ago too, have some benchmarks and tests, happy to work on PR too