Lossless state compression for exact quantum simulation
4.9x

Faster than Google's qsim. Head-to-head, same GPU,
same circuits. Verified.

Up to 34 qubits dense. 37 for structured states. A Recognex product.

Talk to Us See the Benchmarks
Scroll
01 — What They Said

"This cannot fit onto a single consumer 24 GB GPU; a 24 GB GPU natively holds a limit of roughly 30 qubits for exact state-vector simulations."

DenseSpace → 30 qubits dense on A6000. 37 for structured states on 24 GB. 4.9x faster than qsim.
Their Workaround #1
Tensor Networks
Sacrifice exactness. Approximation errors compound over circuit depth.
✕ Not exact
Their Workaround #2
Multi-Node GPU Clusters
Pool 16+ A100/H100 GPUs. $50K+/month infrastructure.
✕ Not accessible
Their Workaround #3
Managed Cloud Services
Offload to BlueQubit-scale infrastructure. No control, no ownership.
✕ Not yours
0
Qubits dense, one GPU
45x
Beyond raw VRAM
39.6x
VQE speedup vs cuQuantum
1.1 TB
State vector, consumer GPU
02 — Proven at Scale

Beyond VRAM.
Beyond belief.

Qubits Raw State GPU Used Method Result
30 8.6 GB 8.6 GB Uncompressed 2.8s gate time
33 68.7 GB ~0.5 GB Proprietary 137x beyond VRAM
34 137 GB H200 150 GB Dense Matches BlueQubit record, 28% less VRAM
35 275 GB 7 GB Sparse chunked 0.37s/gate — 1 of 1024 chunks active
37 1.1 TB 24 GB Proprietary 45x beyond VRAM. Structured states only.

All results independently reproducible. Conservation laws verified to machine epsilon (< 10⁻¹⁴).

03 — Raw Performance

Faster where
it matters.

Gate Throughput vs cuQuantum
2.9x
faster at 15–20 qubits
The practical sweet spot for variational algorithm development. Complex enough that CPU simulation is impractical, small enough for thousands of VQE evaluations.
VQE End-to-End (25 qubits)
39.6x
wall-time reduction
Same Hamiltonian, same hardware, same optimizer. 645 seconds → 16 seconds. The advantage grows with parameter count.

VQE Wall Time — 25 Qubits, 100 Parameters

DenseSpace
16s
cuQuantum
645s
04 — Validated Applications

Real problems.
Real results.

Li
Quantum Chemistry
VQE on lithium hydride (LiH) — 12 qubits, 631 Pauli terms. Full Jordan-Wigner Hamiltonian with hopping terms.
74% correlation energy recovered in 48 seconds on a $0.30/hr GPU
HE
High-Energy Physics
Quantum Born Machine applied to CERN QHack 2025 Challenge — LHC jet transverse momentum prediction from 500K collision events.
35% improvement over baseline → #1 on leaderboard
CO
Combinatorial Optimisation
QAOA on OR-Library portfolio benchmarks at 31 qubits (497 Ising terms). First QAOA result on these standard benchmarks.
31-qubit QAOA on H200 in 7.7 minutes
05 — Competitive Landscape

No one else
does this.

Capability Qiskit Aer Cirq / qsim cuQuantum DenseSpace
GPU-native Partial Yes Yes Yes (exclusive)
Exact time evolution Trotter Trotter Trotter Chebyshev ✓
Beyond-VRAM simulation No No No 45x proven ✓
Max qubits (single GPU) ~30 ~30 34 (H200) 34 dense, 37 structured ✓
VQE speedup (25q) — — Baseline 39.6x ✓
Exploits state structure No No No Yes ✓
06 — Get Started

Lossless compression
for exact simulation.

DenseSpace is a quantum simulation vertical of Recognex. We're working with research groups who need exact classical baselines at scale.

Talk to Us Read the Benchmarks