Static demo · this is a pre-rendered snapshot showing real Akamai Blackwell results from 2026-07-27. Run experiments yourself by cloning the repo. GitHub →
Q
cudaq
molecular simulation blueprint

CPU vs GPU comparison

Aggregated across 24 runs in results/. Per-arm mean ± standard error.

Arms are keyed by backend and ansatz mode where a run declares one. A legacy_full and a matched LiH run share a backend and a Hamiltonian but execute different circuits (12 qubits / 92 parameters versus 10 / 24), so they are never averaged together and CPU/GPU ratios are only computed within a single mode.

H2

0.96x cpu_over_gpu_fp64_wall_time 1.30x cpu_over_gpu_fp32_wall_time
backend n wall time (s) time/eval (ms) evaluations |err vs ref| (Ha)
gpu_fp64 3 17.653 ± 1.076 221.58 ± 0.36 79.7 ± 4.8 1.75e-07
gpu_fp32 3 12.976 ± 0.390 222.43 ± 0.54 58.3 ± 1.7 2.83e-06
cpu 3 16.872 ± 0.833 229.01 ± 1.71 73.7 ± 3.5 1.75e-07

LIH

1.08x matched/cpu_over_gpu_fp64_wall_time
backend n wall time (s) time/eval (ms) evaluations |err vs ref| (Ha)
cpu · matched 5 773.379 ± 30.075 732.91 ± 0.12 1055.2 ± 40.9 2.99e-08
gpu_fp64 · matched 5 719.399 ± 5.503 698.44 ± 0.46 1030.0 ± 7.8 2.99e-08
gpu_fp64 · legacy_full 5 1068.843 ± 3.470 712.56 ± 2.31 1500.0 ± 0.0 3.12e-02