6 min

Production & D-Wave

Real Hardware, Benchmarks & Honest Limitations

From Simulation to Reality

Everything you've built so far runs on classical hardware — the simulated annealing solver uses thermal fluctuations, not quantum effects. In this module, you'll bridge the gap to real quantum hardware and honestly assess what that means.

D-Wave Quantum Annealers

D-Wave Systems builds the only commercially available quantum annealers. Their current processor, Advantage, has:

~5000 qubits arranged in a Pegasus graph topology

~35,000 couplers connecting nearby qubits

~20 microsecond anneal time (the actual quantum computation)

~10 millisecond total per sample (including readout and overhead)

The Embedding Problem

This is the critical constraint that textbooks often gloss over. D-Wave's qubits are not fully connected — each qubit connects to ~15 neighbors in the Pegasus graph. Your QUBO might require connections between any pair of variables.

Minor embedding maps logical variables to chains of physical qubits, where each chain acts as one logical variable. A 30-variable fully-connected QUBO might need 100+ physical qubits after embedding.

This means:

Your effective problem size is much smaller than 5000

Longer chains are less reliable (the chain might break)

Embedding overhead adds noise to the solution

D-Wave Leap

D-Wave offers free access through Leap:

1 minute of QPU time per month (free tier)

That's roughly 6000 samples (at ~10ms each)

Enough for experimentation, not production workloads

https://cloud.dwavesys.com/leap/

Integration Approach

D-Wave's Ocean SDK is Python-based. From TypeScript, you have two options:

Python bridge: Spawn a Python subprocess that submits the QUBO and returns results

REST API: D-Wave's SAPI (Solver API) accepts HTTP requests directly

The QUBO format is the same — you're just changing the backend from your local SA solver to D-Wave's QPU.

Rigorous Benchmarking

A portfolio-worthy project needs benchmarks that answer: "When does quantum optimization help?"

What to Measure

Solution quality: Energy of the best solution found

Consistency: Standard deviation of energy across runs

Time to solution: Wall-clock time to find a solution of given quality

Scaling: How do metrics change as problem size increases?

Problem Sizes to Test

N = 10: Brute force is instant. All methods should find optimal.

N = 20: Brute force takes ~1 second. SA should match.

N = 30: Brute force is slow (~minutes). SA and greedy diverge.

N = 50: Brute force is infeasible. Only heuristics remain.

Methods to Compare

Brute force (N ≤ 20): The ground truth. Guaranteed optimal.

Random selection: No intelligence. The floor.

Greedy: Local intelligence. Fast, decent, can't handle interactions.

Simulated annealing: Global intelligence. Slow but thorough.

D-Wave (if available): Quantum intelligence. Fast per sample, but noisy.

Expected Results (Being Honest)

At the scales in this project (N = 30-50):

SA will likely match or beat D-Wave due to embedding overhead

Both will beat greedy on problems with strong variable interactions

The gap between SA and greedy grows with problem size and interaction strength

D-Wave's advantage (if any) appears at N > 100 where SA struggles with rugged landscapes

Honest Limitations

This is the most important section of your portfolio piece. Intellectual honesty about limitations signals maturity.

No Proven Quantum Advantage (at this scale)

For problems under ~100 variables, well-tuned classical heuristics (SA, tabu search, genetic algorithms) are competitive with or better than current quantum hardware. The overhead of minor embedding and quantum noise offsets any tunneling advantage.

QUBO Formulation is Lossy

Encoding feature selection as MI + correlation + cardinality is an approximation. It doesn't capture:

Non-linear feature interactions

Feature-target relationships beyond pairwise MI

The specific classifier that will use the selected features

Hardware is Noisy

Current quantum annealers have:

Limited qubit connectivity (requires embedding)

Analog control errors (the energy landscape isn't exactly what you specified)

Thermal noise (the system isn't perfectly quantum — it partially thermalizes)

The Framework Has Value

Despite these limitations, the QUBO framework itself is valuable:

It's a principled way to think about combinatorial optimization

It separates problem formulation from solution method

It's future-proof — as quantum hardware improves, the same QUBOs run faster

It generalizes — feature selection, graph partition, scheduling, routing all use the same framework

Portfolio Packaging

A strong portfolio piece includes:

README — Problem statement, approach, results, limitations (all in ~500 words)

Dashboard screenshots — Show the interactive tool in action

Benchmark table — Methods × metrics at multiple problem sizes

"What I Learned" section — The non-obvious insights from doing this work

Code quality — Clean TypeScript, meaningful variable names, comments where the math is non-trivial

The most impressive thing you can show an interviewer isn't "I used quantum computing" — it's "I understood when it helps and when it doesn't."

This is chapter 6 of Quantum Optimization for AI.

Get the full hands-on course for $100 and build the complete system. Your projects become your portfolio.

View course details

Ch. 5: Pipeline App