GPU acceleration via NVIDIA's cuQuantum enables simulation of large quantum systems (20+ qubits) with high performance. The GPU backend exploits parallelism in state vector operations.

#include "cudaq.h"
 
// Large circuit: 25 qubits (requires GPU)
struct LargeCircuit {
  void operator()() __qpu__ {
    int n = 25;
    cudaq::qvector q(n);
 
    // Hadamard layer: O(n) gates
    for (int i = 0; i < n; i++) {
      h(q[i]);
    }
 
    // Entangling layer: two-qubit gates
    for (int i = 0; i < n - 1; i++) {
      cx(q[i], q[i+1]);
    }
 
    // Rotation layer
    for (int i = 0; i < n; i++) {
      rz(0.5 * i, q[i]);
    }
 
    mz(q);
  }
};
 
int main() {
  // Use GPU simulator
  cudaq::set_target("nvidia");
 
  // 25 qubits = 2^25 = 33M complex numbers
  // GPU can handle this; classical simulation would be intractable
 
  printf("Simulating 25-qubit circuit on GPU...\n");
  auto result = cudaq::sample<LargeCircuit>(1000);
 
  printf("Samples collected: %lu\n", result.size());
 
  // Timing: GPU is ~100x faster than CPU for such sizes
 
  return 0;
}

CPU: ~1 μs per gate operation (classical state vector)

GPU (cuQuantum): ~10 ns per gate (parallelized across many cores)

Practical limit: CPU ~20 qubits, GPU ~25–30 qubits (within seconds)

GPU acceleration is critical for testing algorithms before deploying to real hardware.