[QDP] The Scaling Test (Latency vs. Qubits)

[400Ping]

Purpose of PR

• Measure loading time for state sizes ranging from N=10 to N=28 qubits.
• Hypothesis: Circuit-based methods will degrade exponentially (O(N) gates) or linearly (O(N) depth), while QDP will remain effectively constant until PCIe bandwidth saturation.

Related Issues or PRs

Closes #740

Changes Made

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Breaking Changes

• Yes
• No

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes
• Successfully built and ran all unit tests or manual tests locally
• PR title follows "MAHOUT-XXX: Brief Description" format (if related to an issue)
• Code follows ASF guidelines

[400Ping]

cc @rich7420 @guan404ming @ryankert01

[rich7420]

my results:

uv run python benchmark/benchmark_latency.py --qubits 12 --batches 20 --batch-size 8 --prefetch 4
Generating 160 samples of 12 qubits...
  Batch size   : 8
  Vector length: 4096
  Batches      : 20
  Prefetch     : 4
  Frameworks   : pennylane, qiskit-init, qiskit-statevector, mahout
  Generated 160 samples
  PennyLane/Qiskit format: 5.00 MB
  Mahout format: 5.00 MB

======================================================================
DATA-TO-STATE LATENCY BENCHMARK: 12 Qubits, 160 Samples
======================================================================

[PennyLane] Full Pipeline (DataLoader -> GPU)...
  Total Time: 0.0641 s (0.401 ms/vector)

[Qiskit Initialize] Full Pipeline (DataLoader -> GPU)...
  Total Time: 10.5576 s (65.985 ms/vector)

[Qiskit Statevector] Full Pipeline (DataLoader -> GPU)...
  Total Time: 0.0329 s (0.205 ms/vector)

[Mahout] Full Pipeline (DataLoader -> GPU)...
  Total Time: 0.1763 s (1.102 ms/vector)

======================================================================
LATENCY (Lower is Better)
Samples: 160, Qubits: 12
======================================================================
Qiskit Statevector      0.205 ms/vector
PennyLane               0.401 ms/vector
Mahout                  1.102 ms/vector
Qiskit Initialize      65.985 ms/vector
----------------------------------------------------------------------
Speedup vs PennyLane:            0.36x
Speedup vs Qiskit Init:          59.89x
Speedup vs Qiskit Statevec:       0.19x

uv run python benchmark/benchmark_latency.py --qubits 18 --batches 20 --batch-size 32 --prefetch 4
Generating 640 samples of 18 qubits...
  Batch size   : 32
  Vector length: 262144
  Batches      : 20
  Prefetch     : 4
  Frameworks   : pennylane, qiskit-init, qiskit-statevector, mahout
  Generated 640 samples
  PennyLane/Qiskit format: 1280.00 MB
  Mahout format: 1280.00 MB

======================================================================
DATA-TO-STATE LATENCY BENCHMARK: 18 Qubits, 640 Samples
======================================================================

[PennyLane] Full Pipeline (DataLoader -> GPU)...
  Total Time: 3.4523 s (5.394 ms/vector)

[Qiskit Initialize] Full Pipeline (DataLoader -> GPU)...
  Total Time: 227.0904 s (354.829 ms/vector)

[Qiskit Statevector] Full Pipeline (DataLoader -> GPU)...
  Total Time: 3.7708 s (5.892 ms/vector)

[Mahout] Full Pipeline (DataLoader -> GPU)...
  Total Time: 7.3243 s (11.444 ms/vector)

======================================================================
LATENCY (Lower is Better)
Samples: 640, Qubits: 18
======================================================================
PennyLane               5.394 ms/vector
Qiskit Statevector      5.892 ms/vector
Mahout                 11.444 ms/vector
Qiskit Initialize     354.829 ms/vector
----------------------------------------------------------------------
Speedup vs PennyLane:            0.47x
Speedup vs Qiskit Init:          31.01x
Speedup vs Qiskit Statevec:       0.51x

[rich7420]

@400Ping thanks for the patch

[400Ping]

cc @guan404ming PTAL

[guan404ming]

Looks nice!

[guan404ming]

cc @rich7420

[ryankert01]

on colab, any thought of why we are slower then pennylane?

--qubits 12 --batches 20 --batch-size 8 --prefetch 4

Generating 160 samples of 12 qubits...
  Batch size   : 8
  Vector length: 4096
  Batches      : 20
  Prefetch     : 4
  Frameworks   : pennylane, qiskit-init, qiskit-statevector, mahout
  Generated 160 samples
  PennyLane/Qiskit format: 5.00 MB
  Mahout format: 5.00 MB

======================================================================
DATA-TO-STATE LATENCY BENCHMARK: 12 Qubits, 160 Samples
======================================================================

[PennyLane] Full Pipeline (DataLoader -> GPU)...
  Total Time: 0.0689 s (0.431 ms/vector)

[Qiskit Initialize] Full Pipeline (DataLoader -> GPU)...
  Total Time: 19.0670 s (119.169 ms/vector)

[Qiskit Statevector] Full Pipeline (DataLoader -> GPU)...
  Total Time: 0.0230 s (0.144 ms/vector)

[Mahout] Full Pipeline (DataLoader -> GPU)...
  Total Time: 0.2116 s (1.322 ms/vector)

======================================================================
LATENCY (Lower is Better)
Samples: 160, Qubits: 12
======================================================================
Qiskit Statevector      0.144 ms/vector
PennyLane               0.431 ms/vector
Mahout                  1.322 ms/vector
Qiskit Initialize     119.169 ms/vector
----------------------------------------------------------------------
Speedup vs PennyLane:            0.33x
Speedup vs Qiskit Init:          90.13x
Speedup vs Qiskit Statevec:       0.11x