More useful results from quantum hardware.
Prometheus combines hardware-aware compilation, processor selection, and spatial multiplexing to improve execution quality and quantum resource utilization. It works alongside Qiskit SABRE and Quantinuum TKET, with public hardware benchmarks documenting the results.
fez, marrakesh, kingston), Braket (Cepheus-1-108Q) & IQM Resonance (Emerald 54Q, Garnet 20Q)IBM-Heron-156Q-Benchmarks and qpu-selection-execution-quality and are counted once in the deduplicated totals (261 - 4 = 257 jobs; 4,924 - 144 = 4,780 circuits; 11,400,192 - 147,456 = 11,252,736 shots). In spatially multiplexed jobs on IBM Heron and IQM Resonance, multiple independent workloads execute per physical shot, producing 5,790 unique scored workload evaluations and 12,053,504 workload-level shot observations. All multi-window dense-circuit comparisons average repeated windows within each unique logical workload prior to paired nonparametric inference. Simulation and compilation-equivalence checks are reported separately below.
Three Principal Hardware Outcomes
Each outcome is measured on physical quantum processors with a defined cohort, metric, and public dataset.
Dedicated 16-workload timing comparison on ibm_marrakesh (1,024 physical shots/circuit, qualification threshold F_B >= 0.90). Eight-way buffered packing qualified 11/16 workloads in 3 charged seconds (0.5663 s active) versus 9/16 in 7 charged seconds (4.4424 s active) for sequential SABRE O3.
Extended IBM Heron campaign across ibm_fez, ibm_marrakesh, and ibm_kingston (18 logical workloads, 4Q-15Q, 2 blocks = 432 compiled circuits in Pillar 1A of the 1,023-circuit campaign). Prometheus reduced mean two-qubit gates from 122.8 to 108.0 (-12.05%) versus SABRE O3.
Three-window replication on Rigetti Cepheus-1-108Q via Amazon Braket (18 logical workloads, 4Q-17Q, averaged across 3 independent calibration windows; 450 total circuit executions, 460,800 physical shots). Holm-adjusted exact permutation p-values: 0.0112 vs SABRE O3 and 0.000641 vs TKET.
Where Prometheus Fits & Who We Are
Commercial value across the quantum stack and the team leading the execution and molecular workflows.
Hardware Providers UTILIZATION
Evaluate more useful workload output per charged execution second by packing independent circuits onto isolated, high-fidelity subgraphs of utility-scale processors.
Quantum Platform Teams INTEROPERABILITY
Integrate pre-execution compiler selection (evaluating Prometheus, Qiskit SABRE, and TKET candidates), subgraph placement, and multi-QPU dispatch into existing SDK workflows.
Application Researchers FIDELITY
Evaluate whether lower two-qubit gate counts, buffered spatial packing, and calibration-aware processor dispatch translate into more accurate chemistry and optimization outcomes.
William Denyer FOUNDER · PROMETHEUS DYNAMICS
Architect of the Prometheus execution management, compilation, and spatial-multiplexing stack. Leads multi-cloud hardware validation across IBM Quantum Heron processors and Amazon Braket superconducting backends.
Dr. Debasis Sahu, Ph.D. CO-FOUNDER · QUANTOMICS / QUANOVIS
Scientific co-founder leading macromolecular dynamics, 4D conformational screening, and therapeutic candidate prioritization. Together, Prometheus Dynamics and Quanovis develop QuantOmics, linking classical candidate triage with active-space quantum chemistry.
Hardware Benchmark Summary
Concise campaign overview (zero scrollbars), dedicated timing comparison, and expandable full ledgers and figures.
| Hardware Campaign & Repository | QPUs · Jobs · Circuits · Shots | Measured Output Quality (↑) | Mean 2Q Gates / Depth (↓) | Cohort Comparison |
|---|---|---|---|---|
|
IQM Emerald Dense Optimization & Matched Placement
48 dense (7Q-12Q) · 16 matched-placement · 12 high-width (13Q-17Q)
|
IQM Emerald 54Q19 jobs · 739 circuits · 756,736 shots |
0.7389 F_B (N=48 wk mean) SABRE K8: 0.7123 · Pure TKET K8: 0.6869 Matched Placement: 0.7222 vs 0.6852 SABRE K8 |
139.4 CZ gates (-8.3% vs SABRE K8, -44.5% vs Pure TKET K8) 2Q Depth: 54.4 vs 81.6 (-33.3%) |
+3.73% vs SABRE K8 · +7.57% vs Pure TKET K8 +5.40% matched placement (16/16 wins) · +9.37% 13Q-17Q (12/12 wins) |
|
IBM Heron Extended Validation
Pillar 1A (432 circuits) · 1,023 total
|
3 IBM Heron QPUs 17 jobs · 1,023 circuits · 1,047,552 shots |
0.5761 F_B (mean) SABRE 0.5578 · TKET 0.4963 |
108.0 2Q gates (mean) SABRE 122.8 · TKET 184.6 Depth: -14.19% vs SABRE |
+3.27% F_B vs SABRE +5.42% F_H (+6.67% dispatch) |
|
Rigetti 3-Window Replication
18 workloads x 3 windows
|
Rigetti Cepheus-1-108Q10 tasks · 450 circuits · 460,800 shots |
0.6542 F_B (mean) SABRE 0.6296 · TKET 0.6001 |
32.78 2Q gates (mean) SABRE 37.83 · TKET 43.22 Depth: -23.56% vs SABRE |
+3.91% F_B vs SABRE p_Holm = 0.0112 (+9.03% TKET) |
|
IQM Resonance Multiplexing & Scheduler
2-Window 60-Wk · Stage 3 Delay · Stage 4A/4B Scheduler
|
IQM Emerald 54Q & Garnet 20Q40 jobs · 740 PUBs (1,196 evals) · 378,880 shots |
Prom 1W: 0.9754 F_B · 2W: 0.9741 · 4W: 0.9705 SABRE Def: 0.9662 · TKET Def: 0.9571 Stage 4B Stress Scheduler: 0.9458 vs 0.9078 4W |
20.48 2Q gates (60-wk mean) SABRE 21.45 · TKET 22.50 Stage 3: ~78.1% of 4W drop reproduced by delay |
1.86x (2W) & 3.30x (4W) qual/s 1.67x (2W) & 3.00x (4W) qual/credit (p = 0.00092) |
|
Compiler Interoperability (8-Way)
Campaign 3 · 16 workloads x 3 QPUs
|
3 IBM Heron QPUs 19 jobs · 586 PUBs · 655,360 shots |
0.8796 F_B (Meta-Selector) Prom 0.8777 · SABRE 0.8776 · TKET 0.8761 |
62 active qubits / PUB Disjoint subgraphs (d_G >= 2) |
TKET packed +0.0031 F_B vs isolated TKET (p = 0.0117) |
|
Heron 156Q Odd/Even & 15-Family
99 tournaments + 18 pre-committed
|
3 IBM Heron QPUs 10 jobs · 441 circuits · 755,712 shots |
0.8494 F_B (99-tourn mean) SABRE 0.8447 · TKET 0.7291 |
Depth <= SABRE on 84/99 100% (54/54) in 3Q-13Q sweep |
16/18 pre-committed wins 65/99 wins vs SABRE · 89/99 TKET |
|
Multi-Cloud 3-Way Corpus
107 Heron jobs + 12 Braket tasks
|
6 Physical QPUs 119 jobs · 357 circuits · 5,084,160 shots |
74.8% win rate vs SABRE 93.5% win rate vs TKET (107 Heron) |
256.6 mean depth (Heron) SABRE 321.2 · TKET 464.1 |
9/12 Braket fidelity wins 12/12 Braket depth reductions |
| Dedicated Timing & Billing Comparisons (Qualification Threshold F_B >= 0.90) | Mean Overlap F_B (↑) | Qualifying Workloads (F_B >= 0.90) (↑) | Active QPU Time (s) / Billed Cost (↓) | Qualifying Output Rate (↑) |
|---|---|---|---|---|
| IBM Marrakesh (156Q): SABRE O3 1-Way Sequential (16 PUBs, 1,024 shots, T1) | 0.8593 (mean) | 9 of 16 workloads | 4.4424 s active · 7 s billed | 1.29 qual / billed s |
| IBM Marrakesh (156Q): Prometheus 8-Way Buffered (2 PUBs, 1,024 shots, T4) | 0.8837 (mean) | 11 of 16 workloads | 0.5663 s active · 3 s billed | 3.67 qual/billed s (2.85x) · -57.1% billed |
| IQM Emerald (54Q): Default SABRE O3 1-Way (Pooled 2 Windows) (120 PUBs, 120 evals, 512 shots) | 0.9662 (mean) · 0.9757 (median) | 116 of 120 workloads (96.7%) | 27.0538 s active · 22.50 credits (30 s) | 4.29 qual/active s · 5.16 qual/credit |
| IQM Emerald (54Q): Prometheus Buffered 1-Way (Pooled 2 Windows) (120 PUBs, 120 evals, 512 shots) | 0.9754 (mean) · 0.9805 (median) | 120 of 120 workloads (100%) | 27.5702 s active · 22.50 credits (30 s) | 4.35 qual/active s · 5.33 qual/credit (1.00x) |
| IQM Emerald (54Q): Prometheus 2-Way Packed (Pooled 2 Windows) (60 PUBs, identical gates/qubits) | 0.9741 (mean) · 0.9789 (median) | 120 of 120 workloads (100%) | 14.8113 s active · 13.50 credits (18 s) | 8.10 qual/active s (1.86x) · 8.89 qual/cr (1.67x) |
| IQM Emerald (54Q): Prometheus 4-Way Packed (Pooled 2 Windows) (30 PUBs, identical gates/qubits) | 0.9705 (mean) · 0.9778 (median) | 120 of 120 workloads (100%) | 8.3584 s active · 7.50 credits (10 s) | 14.36 qual/active s (3.30x) · 16.00 qual/cr (3.00x) |
| IQM Emerald (54Q): Stage 4B Boundary-Stress Prospective Scheduler (16 5Q workloads: 12 PUBs 2+1+1 vs 4 PUBs 4W) | 0.9458 (Sched) vs 0.9078 (Always 4W) | 14/16 >= 0.88 (Trotter 0.8793 vs 0.7863) | 3.5230 s active · 3.00 credits (-20% vs 1W) | +0.0379 vs 4W (recovers 99.0% of 4W loss; 99.96% of 1W) |
| IQM Garnet (20Q): Matched 1-Way vs. 2-Way Contemp. (T3 8 PUBs vs T4 4 PUBs) | 0.9676 (1W) vs 0.9630 (2W) | 7 of 8 workloads (both) | 1.6174 s vs 1.1029 s active (-31.8%) · 1.00 cr each | 6.35 vs 4.33 qual/active s (1.47x) |
Expand Representative Single-Job Hardware Benchmarks & Separate Simulation / Compilation Studies [Click to Expand / Collapse]
Part A: Physical QPU Single-Job Benchmarks (measured on superconducting hardware via AWS Braket ProgramSet and IBM Quantum SamplerV2):
| Processor & Hardware Job / Task ID | Workload | Compiled Depth (SABRE / TKET / Prom) (↓) | Compiled 2Q Gates (SABRE / TKET / Prom) (↓) | Measured Hardware Overlap F_B (↑) |
|---|---|---|---|---|
| Rigetti Cepheus-1-108Q (2df64966-8b30) | 9Q Dense All-to-All | 137 / 131 / 63 (-54.0%) | 80 / 83 / 54 (-32.5%) | 0.2088 F_B (vs 0.0435 SABRE, 0.0211 TKET) |
| IQM Emerald 54Q (68dbffa0-f842) | 6Q Quantum Fourier Transform | 88 / 89 / 64 (-27.3%) | 44 / 48 / 38 (-13.6%) | 0.9602 F_B (vs 0.8355 SABRE, 0.9648 TKET) |
| IQM Garnet 20Q (8a04e33b-e945) | 9Q Dense All-to-All | 130 / 129 / 103 (-20.8%) | 79 / 87 / 70 (-11.4%) | 0.6031 F_B (vs 0.5513 SABRE · HOP 0.776) |
| ibm_kingston 156Q (dbajlrhgv2cc73e6vuj0) | 19Q Dense All-to-All | 358 / 501 / 281 (-21.5%) | 444 / 717 / 358 (-19.4%) | 0.2426 F_B (vs 0.1898 SABRE, 0.0162 TKET) |
| ibm_marrakesh 156Q (dbaijj9gv2cc73e6ubmg) | 8Q Quantum Fourier Transform | 226 / 288 / 175 (-22.6%) | 104 / 137 / 85 (-18.3%) | 0.9929 F_B (vs 0.9862 SABRE, 0.9579 TKET) |
Part B: Compilation & Classical/GPU Simulation Benchmarks (evaluated via exact statevector equivalence or GPU solver; separate from QPU shot totals):
| Target Architecture / Environment | Evaluation Type & Workload | Structural Reduction (↓) | Verification Outcome |
|---|---|---|---|
| IonQ Forte 1 (36Q) / AQT IBEX Q1 (12Q) | Native all-to-all RZZ / RXX KAK compilation (6Q QFT) | Depth: 226 / 223 / 56 (-75.2%) · 2Q gates: 39 to 15 (-61.5%) | Exact unitary statevector equivalence (1.0000); -79.2% 1Q laser pulses |
| NVIDIA H100 GPU (VeloxQ) + QUBO Reduction | 42 scheduling configs (72-var to 1,029-var binary QUBO) | Algebraic presolve: 72 to 4 binary variables (-94.44%) | 1,029-variable schedule solved in 5.26 s on NVIDIA H100 (0.0% optimality gap) |
Application Section: QuantOmics & Active-Space Quantum Chemistry
Led with Dr. Debasis Sahu, Ph.D., connecting blinded macromolecular candidate prioritization with active-space quantum chemistry (LUCJ-SQD).
Blinded Candidate Prioritization RECONCILED ENERGY RANKS
Validated on 52 experimental PDB complexes (82.7% [43/52] positive concordance, Kendall tau_b = 0.258, p_Holm = 2.09e-7). Candidates flagged HOLD by stereochemical (306-center) or trajectory stability gates are excluded prior to multi-tier ranking. Among the 4 PASS candidates below, consensus rank is the unweighted mean of within-method ordinal ranks:
| Blinded ID | OBC2 Mean (kcal/mol) (↓) | GBn2 Mean (kcal/mol) (↓) | GFN2-xTB Pocket Delta E (↓) | Consensus Rank |
|---|---|---|---|---|
| Candidate 07 | -53.62 (#1) | -43.65 (#1) | -69.29 (#1) | #1 (Mean 1.00) |
| Candidate 03 | -39.63 (#2) | -36.71 (#2) | -42.10 (#2) | #2 (Mean 2.00) |
| Candidate 06 | -29.16 (#4) | -26.23 (#3) | -32.94 (#3) | #3 (Mean 3.33) |
| Candidate 11 | -35.49 (#3) | -26.14 (#4) | -30.09 (#4) | #4 (Mean 3.67) |
Method Note: Candidate 07 ranks first across all three energy tiers, while Candidates 06 and 11 invert between #3 and #4 across OBC2 versus GBn2 and GFN2-xTB. Binding (SPR/BLI K_d) and efficacy (Vero-cell EC_50) assays are the wet-lab validation endpoints.
Quantum Chemistry Handoff (LUCJ-SQD) SIMULATOR CHECK -> QPU PILOT
QuantOmics exports active-space Hamiltonians for Sample-Based Quantum Diagonalization (LUCJ-SQD). Our completed 6-qubit handoff (bundle sqd-29cc656768a9) is a statevector simulator integration check on a Candidate 07 active-site fragment:
| Stage / Reference | Scope & Circuit Structure | Energy Attribution & Test Goal |
|---|---|---|
| Exact CASCI Reference 6Q (2e, 3o) STO-3G |
Classical exact active-space diagonalization | -968.53388903 Ha (exact within-model reference) |
| HF & LUCJ-SQD Check Bundle sqd-29cc656768a9 |
6Q logical circuit (32 2Q gates, depth 40, 4,096 simulator shots) | -968.53388101 Ha (both HF and SQD; 0.008 mHa / ~0.005 kcal/mol from CASCI) |
| Proposed 36Q-56Q QPU Models Hardware Value Test |
Larger correlated active spaces on IBM Heron & Braket QPUs | Measure chemistry error and usable samples/s where correlation exceeds HF (packing scaled to width) |
Topology Scaling Note: Packing density scales with circuit width; for example, eight 36-qubit circuits require 288 active qubits before guard buffers and cannot be packed 8-way on a 156-qubit processor.
Coordinated Multi-QPU Dispatch & Multiplexing Evaluation
Queue delays increase the age of calibration data, allowing hardware drift or recalibration to alter which subgraphs and processors perform best. We propose a concrete technical pilot with coordinated access to at least three QPUs in overlapping windows to measure:
- Calibration-to-Execution Delay & Drift: Record calibration snapshot, compilation, submission, and execution timestamps across short-delay and delayed controls to quantify queue-related drift alongside prediction error.
- Live Compiler & Processor Selection: Evaluate Prometheus, Qiskit SABRE O3, and Quantinuum TKET across at least 3 QPUs under comparable timing.
- Multiplexed Throughput & Chemistry Scaling: Measure qualifying workloads per charged QPU second and active-space chemistry error from 6Q-12Q fragments up to 36Q-56Q correlated models.
Nine Public GitHub Evidence Repositories
Complete per-repository accounting of physical QPUs tested, jobs/tasks, compiled circuits/PUBs, and physical shots (4 shared pre-committed Heron jobs / 144 circuits / 147,456 shots are archived in both IBM-Heron-156Q-Benchmarks and qpu-selection-execution-quality and counted once in the 257-job / 4,780-circuit / 11.25M-shot totals).