Proven Performance. Unmatched Efficiency.
Verified benchmarks on NVIDIA GPUs. 5,907x more cost-efficient than the nearest commercial competitor.
Verified benchmarks on NVIDIA GPUs. 5,907x more cost-efficient than the nearest commercial competitor.
Benchmarks run on two NVIDIA GPUs spanning consumer and workstation tiers.
Consumer — $600 MSRP
Workstation — $6,800 MSRP
DSC-2 is a member of the NVIDIA Inception Program — accelerating innovation with NVIDIA GPU computing.
Millions to billions of spin-flips per second across dense, sparse, and batched workloads. Every chart shows DSC-2 running on commodity NVIDIA hardware.
Full-density coupling matrices — both GPUs overlaid. Higher is better.
Sparse random graphs with degree k=6, up to 30M variables. Log scale on x-axis.
Concurrent batch execution — peak 1.67 billion spins/s at N=5K × 100 batch.
DSC-2 accuracy against Best Known Solutions on standard G-set benchmark graphs.
| Instance | Variables | Best Known | DSC-2 (5070 Ti) | Accuracy | DSC-2 (6000 Ada) | Accuracy |
|---|---|---|---|---|---|---|
| G1 PERFECT MATCH | 800 | 11,624 | 11,624 | 100.00% | 11,624 | 100.00% |
| G14 | 800 | 3,064 | 3,057 | 99.77% | 3,061 | 99.90% |
| G22 | 2,000 | 13,359 | 13,183 | 98.68% | 13,290 | 99.48% |
| G43 | 1,000 | 6,660 | 6,639 | 99.68% | 6,652 | 99.88% |
| G55 | 5,000 | 10,294 | 10,116 | 98.27% | 10,200 | 99.09% |
On the RTX 6000 Ada, DSC-2 achieves 99.09% to 100% accuracy across all tested G-set instances. On the $600 RTX 5070 Ti, accuracy ranges from 98.27% to 100%. These results are achieved in milliseconds, not hours.
BKS values from published literature. All DSC-2 results from 10-run best-of with 1000 sweeps.
DSC-2 on a $600 consumer GPU outperforms million-dollar quantum annealers by orders of magnitude.
DSC-2 delivers 5,907x more cost-efficient optimization than the nearest commercial competitor.
How DSC-2 compares across every category of optimization platform.
D-Wave, Toshiba, Fujitsu
CPLEX, Gurobi, Concorde
IBM Quantum, Braket, Azure Quantum
DSC-2 is purpose-built for Ising/QUBO-native problems: MaxCut, graph partitioning, scheduling, portfolio optimization, molecular conformation, network design, and resource allocation.