Hi NVIDIA team,
Four days after the NVIDIA Ising launch on April 16, I built and
open-sourced what appears to be the first external downstream
application of Ising Calibration — applied to a new domain beyond
its original QPU calibration design.
Project: GPU-Accelerated Ising Optimisation for Portfolio Selection
GitHub: GitHub - maddykws/quantum-portfolio-ising-claude: Quantum portfolio optimisation: NVIDIA Ising + CUDA-Q + Claude AI. 373x GPU speedup · +3.1% vs classical · 1.4M× sampling efficiency · QPU validated on Rigetti Cepheus-1-108Q. · GitHub
Contact: https://www.linkedin.com/in/cloudaiengineer/
Core results
-
373× GPU speedup — NVIDIA L4 vs CPU at N=20 qubits (CUDA-Q nvidia vs cpu target, identical kernel code)
-
1.4 million × sampling efficiency — median 3 shots to near-optimal portfolio vs 1.4M random draws across the 3.2M combination space
-
+3.1% median Sharpe — vs top-N equal weight across 56 quarterly S&P 500 periods (2010–2024), 75% win rate
-
−4.8% vs top-N optimal weight — disclosed honestly as the hard classical baseline
-
QPU validated — Rigetti Cepheus-1-108Q via Amazon Braket, parameters transferred directly from L4 GPU simulation ($1.44 total)
Ising Calibration application (new domain)
NVIDIA Ising Calibration is designed for QPU hardware calibration.
I applied it to a different domain: assessing QAOA measurement
distribution quality in portfolio optimisation circuits.
Evaluated three pipeline configurations using NVIDIA’s QCalEval
benchmark scoring framework applied to portfolio circuit outputs
across 10 windows:
| Pipeline | Circuit quality | Financial insight | Overall |
|---|---|---|---|
| Ising Calibration alone | 3.9 / 5 | 1.0 / 5 | 2.42 / 5 |
| LLM orchestration alone | 1.0 / 5 | 3.9 / 5 | 2.62 / 5 |
| Combined | 3.9 / 5 | 4.0 / 5 | 2.70 / 5 |
Finding: Ising Calibration’s circuit interpretation capability
transfers to this new domain. It correctly identifies low measurement
concentration as poor QAOA convergence. An LLM orchestration layer
provides financial translation Ising does not attempt. Non-overlapping
capabilities confirmed.
Honest framing
This is a hybrid quantum-classical workflow validation, not a
quantum advantage claim. Ising Calibration applied outside its
intended QPU calibration domain is an early domain extension
experiment, not validated production use.
Reproducibility
Full pipeline runs end-to-end on Google Colab L4 GPU for
approximately $3. One-click Colab badge in README. Two free API
keys required (NVIDIA NIM and Anthropic).
Would the team be interested in featuring this as a Spotlight post?
Happy to provide the full writeup, figures, benchmarks, or adjust
framing based on editorial feedback.
Thanks,
maddykws