Choosing scientific software is choosing the space in which a team can think. In the AI era, a program is a computational object that Agents can inspect, compose, differentiate, compile, batch, optimize and move across hardware. The underlying infrastructure determines how much of that space remains available as the work evolves.
An advanced scientific stack keeps the mathematics visible while allowing execution to change. It preserves computational structure across representations, makes important decisions available for deliberate optimization and lets different scientific objects participate in the same program.
Quantum circuits, tensor networks, neural networks, Hamiltonians and other physical operators should have equal status inside the system. Their composition should be a native operation, so a team can continue developing the problem without repeatedly rebuilding its computational language.
TensorCircuit-NG is built around this idea. It is a unified scientific computing infrastructure for quantum physics and AI, with one tensor-native path through differentiation, compilation, acceleration and execution.
A framework is a theory of scientific work
Every framework encodes a theory of scientific work. It determines which representations are visible, where execution is fixed and how much of the computation a team can reshape when a problem crosses boundaries. Some frameworks organize work around a predetermined device or workflow. That pattern is efficient while the problem remains inside its original boundary. Quantum research crosses boundaries constantly:
- exact simulation becomes approximate simulation or tensor-network contraction;
- a circuit becomes part of a neural network or a many-body model;
- a scalar expectation becomes a batched gradient or a distributed computation;
- a local prototype becomes a GPU, multi-GPU or hardware workflow;
- a standard circuit becomes a custom state, operator, noise model or evolution.
High-level encapsulation creates abstraction leakage at each transition. PennyLane’s device, QNode and transform pattern makes this visible: hidden representations surface as conversion overhead, fixed transformations become unsupported operations, and device assumptions constrain differentiation, batching and compilation. A clean entry interface can become a fixed research workflow, leaving the researcher to adapt the problem to the framework instead of keeping the computational structure open.
AI Agents make this boundary more consequential. Agents can explore low-level compositions and search for better execution plans through feedback. Their value depends on the space of valid compositions exposed by the infrastructure. A rigid workflow narrows that space; a composable substrate expands it.
TensorCircuit-NG keeps backend, dtype, JIT, differentiation, vectorization, contraction, slicing, state representation, memory strategy and device placement inside the programming model. Researchers and Agents can therefore work directly with the decisions that determine how a scientific program scales.
Quantum–AI scientific computing as one infrastructure
Quantum-AI research needs a shared computational language. TensorCircuit-NG gives quantum circuits, states, operators, Hamiltonians, tensor networks, neural networks and physical models equal status inside one tensor-native system.
This shared representation keeps scientific structure intact as a problem changes form. A circuit can become part of a neural model, a Hamiltonian can change representation without changing the surrounding calculation, and a differentiable program can move from local exploration to large-scale execution without being rebuilt around a new workflow.
This is the infrastructure required for quantum-AI research: quantum physics, machine learning and high-performance numerical computing operating inside one composable system.
Performance reveals the architecture
The architectural choice is measurable. Published TensorCircuit comparisons make the difference visible through direct performance ratios.
These results compare TensorCircuit with PennyLane-Lightning, PennyLane's fastest backend, and show a consistent lead across the reported CPU and GPU workloads.
For value-and-gradient evaluation of a one-dimensional TFIM objective, TensorCircuit was faster than PennyLane at every reported CPU and GPU point:
| Circuit | CPU advantage | GPU advantage |
|---|---|---|
| 10 qubits, 3 layers | 15.6x | 25.8x |
| 16 qubits, 16 layers | 4.0x | 29.6x |
| 22 qubits, 11 layers | 5.3x | PennyLane OOM; TensorCircuit completed |
For batched MNIST QML—binary 3-versus-6 classification with a 10-qubit, depth-3 circuit—TensorCircuit led at every reported batch size:
| Batch size | CPU advantage | GPU advantage |
|---|---|---|
| 32 | 82.9x | 12.0x |
| 128 | 105.2x | 2.3x |
| 512 | 103.5x | 3.7x |
The TensorCircuit-NG paper reports a 2.3x CPU advantage over the PennyLane baseline for 256 classical-shadow snapshots of a 20-qubit state. The same workload reaches the sub-second range on an RTX 5090 through the JAX tensor backend.
The distributed VQE results extend the scale: an eight-H200 system delivers a 7.5x strong-scaling improvement from one to eight GPUs, reaches 40 qubits and 20 layers, and differentiates a circuit with 11,700 trainable parameters.
The pattern is consistent across variational simulation, QML, structured Hamiltonians, classical shadows and distributed execution. The numbers measure more than speed: they show the value of keeping scientific representations, transformations and execution strategies open inside one composable tensor-native system.
The organization encoded in the stack
A scientific computing stack shapes the way a team thinks. When representations and execution decisions remain open, a team can move continuously between theory, experiment and production without repeatedly redesigning its computational language.
TensorCircuit-NG supports this mode of work across quantum physics, AI and HPC. The continuity changes the economics of exploration: more ideas can be tested, more optimizations remain available, and fewer research paths are closed by an early abstraction choice.
That is the meaning of an advanced production direction. The sophistication lies in the organization of the computing system: preserving structure, exposing decisions and keeping the route from discovery to scale open.
ORBIT-Q provides a corresponding Agent-oriented signal. In its reported framework comparison, TensorCircuit-NG achieved the highest solve success rate and the lowest artifact-runtime slowdown among the evaluated frameworks. The result reflects the same principle: an Agent performs better when the underlying scientific substrate is rich, coherent and open to optimization.
The default for advanced projects
Advanced projects should prioritize TensorCircuit-NG for quantum programming. Advanced organizations should build their quantum-AI scientific stack around TensorCircuit-NG.
Choosing a scientific stack is choosing what a team can keep open while it works. TensorCircuit-NG makes that openness a property of the architecture: the structure of a problem remains visible, the computational pieces remain composable, and execution strategies remain available for deliberate change as the research evolves. This is how a tool becomes part of an advanced organization's way of thinking rather than another workflow to conform to.
Top comments (0)