BlueQubit has released bluesv, a new exact state vector simulator built entirely without human coders. Claude wrote every line of bluesv’s code, working in Claude Code, after BlueQubit provided detailed design documents and structured build instructions. The simulator is application programming interface-compatible with NVIDIA’s cuStateVec, allowing smooth integration with existing quantum computing tools.

On GPUs, BlueQubit benchmarked randomly generated circuits from 22 to 30 qubits and depths from 10 to 100 and found bluesv is faster at every point measured, over 1.4 times faster on the largest, hardest circuits, the company says. On CPUs, BlueQubit ran the same kind of benchmark and found bluesv is faster at every size tested, about 1.5-1.6x faster at depth 10 and about 1.8x at depth 30 from 26 qubits up.

bluesv Performance Benchmarks Against cuStateVec

bluesv demonstrably outperforms NVIDIA’s cuStateVec on standard benchmarks, achieving over 1.4 times faster computation on circuits with up to 30 qubits and a depth of 100 when tested on an NVIDIA A100 GPU. This advantage extends to CPU performance as well, with bluesv proving approximately 1.5 to 1.6 times faster than cuStateVec’s CPU backend on circuits of 26 qubits and a depth of 10, according to BlueQubit. At a depth of 30 qubits, the performance gap widens to around 1.8 times faster, suggesting hardware limitations are the primary constraint at these scales.

Unlike traditional software development, where code correctness is often assumed, bluesv was built against a deterministic “oracle”, NVIDIA’s cuStateVec, allowing for immediate and precise validation of each code component. Every element of bluesv was tested and refined until it aligned with cuStateVec’s output, as well as a NumPy reference implementation, in both single and double precision.

This approach ensured that any discrepancies were identified and resolved, guaranteeing the reliability of the simulator. The core computational engine within bluesv is inherently faster than its NVIDIA counterpart, according to BlueQubit, and the company reports that the simulator matches cuStateVec and qsim to machine precision on all tested operations.

Beyond passing internal tests and the qsimcirq test suite, bluesv’s performance suggests a new approach to building complex scientific software is viable. This agentic engineering approach reduced development time from an expected year to just weeks, freeing up BlueQubit’s team to focus on algorithm development and experimentation. BlueQubit stated, “The bigger point is about how this field gets built,” and that “AI can now write advanced, domain-specific libraries that hold their own against hand-tuned originals, as long as the design is careful, the specifications are clear, and every step can be checked.”

bluesv Verification Process Using Deterministic Oracle

This enabled BlueQubit to automate much of the testing process; Claude Code, operating from the detailed design documents, repeatedly wrote code kernels, verified them against the oracle, identified disagreements, and implemented corrections, a cycle repeated thousands of times. The team found that the agentic approach focused on automating the entire inner loop of software development, the company says.

The most significant differences observed occurred at the limits of floating-point precision, which highlights the meticulous verification process. BlueQubit’s method, however, is not universally applicable. The company acknowledges that its success depends on the availability of a “cheap, exhaustive way to check your work.” If a problem lacks a readily available oracle for verification, this agentic engineering approach cannot be directly transferred. The resulting library now powers state-vector simulations within the BlueQubit SDK and app. bluequbit. io, offering users access to large-scale simulations without the need for individual hardware provisioning or software configuration.

Source: https://www.bluequbit.io/blog/bluesv-state-vector-simulator

More like this