China once again raises the bar in supercomputing achievement. A research team has demonstrated on the new supercomputer Sunway “OceanLite” that neural networks of quantum states (NNQS) can scale to relevant molecular sizes, with near-perfect efficiency and across tens of millions of cores. This isn’t just a performance feat: it suggests a pragmatic pathway to model materials and chemical reactions using the most powerful classical computing available today, without waiting for functional quantum processors.
What exactly have they done
The work extends an approach already known in the literature — NNQS — to exa-scale, which trains a network to approximate the wavefunction of an electronic system and thereby predict properties and energies. This training involves two very costly steps: sampling configurations (like “photos” of the quantum system) and calculating the local energy for each. As molecules grow larger, these steps become unbalanced and the calculation becomes irregular; scaling it efficiently is the real challenge.
On the Sunway system, the authors deployed a NNQS-Transformer tailored to the architecture: lightweight cores that compute within local memory (“scratchpad”) and management cores overseeing communications, with a load balancing designed so that all cores work constantly. Result: the code ran on approximately 37 million cores with 92% strong scaling and 98% weak scaling— meaning almost no loss in efficiency when adding resources to the same problem or when increasing the problem size with more resources. Such efficiencies at this scale are rare in quantum chemistry.
Why does it matter
Until now, NNQS often remained confined to small systems. This demonstration has achieved structures of up to 120 spin orbitals, bringing the method closer to molecules and materials of practical interest. If these algorithms maintain such performance at large scale, exa-scale supercomputers could accelerate discovery of new compounds, catalysts, or drugs well before mature quantum computers are available for these tasks.
Additionally, the experiment reinforces a foundational insight: architectures used for training LLMs today can also learn the structure of matter. That a system designed for regular loads (deep learning training) can solve an irregular load (quantum chemistry) with such efficiencies speaks to software maturity, data partitioning, and fine communication and I/O work.
Inside Sunway: the machine profile
Sunway “OceanLite” is the successor to TaihuLight and uses SW26010-Pro chips, organized into clusters with close memory (“scratchpad”) instead of a conventional cache hierarchy. Tens of thousands of these chips interconnect into a system of over 40 million cores, demonstrating exa-scale performance across various scientific challenges. Its data locality and fine memory control fit well with a NNQS-Transformer approach optimized for this topology.
An interesting detail is the use of Julia in the software stack: a high-level language that, combined with highly optimized specific code, allows for rapid prototyping without sacrificing performance close to C/Fortran. A few years ago, this would have sounded experimental; today it’s a staple in productive HPC environments.
Current limits (right now)
Scaling the mathematics was the first barrier; moving and managing data will be the second. These simulations generate massive volumes and require millions of cores to share information constantly. The bottleneck is shifting to storage and network, which must deliver data nearly as fast as computation. The team themselves point to this frontier: without adequate I/O, theoretical efficiency remains on paper.
Furthermore, expanding the system repertoire is necessary: beyond 120 spin orbitals, testing with complex geometries, excited states, and environments (solvents, surfaces) is needed. The goal isn’t just scale, but chemical accuracy compared to reference methods (like cluster coupling) in strongly correlated regimes.
Implications for AI-science collaboration
The message to the ecosystem is twofold. For chemistry and materials science, AI doesn’t replace quantum physics; it approximates it with an increasingly attractive accuracy/cost ratio on modern supercomputers. For HPC, it shows that the exa-scale path isn’t just about benchmarks but about opening new science through aligned algorithms, architectures, and software.
In the medium term, expect collaborative efforts: chemistry groups providing validation sets and metrics; HPC teams refining communication, heterogeneous programming, and storage; and memory/network providers offering topologies more NNQS-ready. If these collaborations solidify, materials and drug discovery could advance sooner than anticipated.
Context and background
The idea of employing neural networks for wavefunctions isn’t new, but its scale leap is. The community has already proposed Transformer variants for NNQS with good metrics and scalability on medium problems; likewise, the Sunway team has demonstrated massive quantum simulations (e.g., random circuits) with near-ideal efficiency. What differentiates now is combining both— NNQS + Sunway— and demonstrating regularity where before there was irregularity and bottlenecks.
What’s next
- More real chemistry: validation with larger datasets and direct comparisons against standard ab initio methods.
- Next-gen I/O: smarter networks and faster storage so computation doesn’t wait on data.
- Portability: reproducing these results on other exa-scale architectures (U.S., EU, Japan) to establish the approach as a working tool, not just a milestone.
via: vastdata and IEEE Explore