| Benchmark | Classical Baseline (GPU) | JUQ‑253 (Hybrid) | Speed‑up | Power (W) | |-----------|--------------------------|------------------|----------|-----------| | | 1.8 s (RTX 4090) | 0.27 s | 6.7× | 210 W | | Variational Quantum Classifier (VQC) – MNIST | 45 ms (GPU) | 12 ms | 3.8× | 230 W | | Monte‑Carlo Option Pricing (10⁶ paths) | 0.62 s (CPU) | 0.14 s | 4.4× | 190 W | | Quantum‑Secure Key Exchange (1 TB) | 18 h (software TLS) | 5 h (QKD off‑load) | 3.6× | 250 W |
– The ARM CPU orchestrates data movement, the GPU handles classic tensor operations, and the QPU is called via a low‑latency PCIe link. The three layers share a unified memory space (256 GB HBM2E) so data never has to leave the board for quantum acceleration. juq-253