Molecular dynamics (MD) simulations are crucial across numerous scientific disciplines, yet face a persistent trade-off between accuracy and efficiency. AIMD, while highly accurate, suffers from computational limitations, restricting its use to small systems. CMD, based on classical force fields, offers high efficiency but compromises accuracy. Machine learning (ML) approaches, such as MLMD, attempt to bridge this gap, achieving AIMD-level accuracy with increased efficiency compared to AIMD. However, even the best MLMD methods remain significantly slower than CMD. This paper addresses this challenge by proposing a new MD methodology that combines high accuracy with CMD-level efficiency. The significant limitation of traditional von Neumann architecture computers, which suffer from data shuttling bottlenecks, is tackled by employing a specialized non von Neumann architecture. This architecture is optimized for the specific operations required for MD calculations, thus mitigating the data transfer overhead and improving overall computational efficiency.

The paper reviews existing MD methods, highlighting the limitations of AIMD and CMD. It discusses the advancements made by MLMD in improving accuracy while increasing efficiency compared to AIMD, but notes that MLMD is still considerably slower than CMD. The authors emphasize the memory wall bottleneck inherent in von Neumann architectures and how this is especially problematic for MD simulations due to the repeated data shuttling required. The lack of accurate and efficient special-purpose MD computers beyond the von Neumann paradigm is identified as a critical gap in the field. Existing special-purpose MD computers, while efficient, typically rely on CMD and therefore lack the accuracy needed for many applications.

The proposed methodology leverages DeepMD, a machine learning model, for high-accuracy PES representation. A key innovation is the deployment of a modified DeepMD algorithm, specifically designed for the non von Neumann architecture. This involves several modifications: quantization of the neural network (QNN) to reduce computational complexity, replacement of multiplication operations with bitwise shifts to minimize hardware resource usage, and the design of a new lightweight activation function without trigonometric functions to improve efficiency. The overall system architecture is heterogeneous, comprising a master processing unit (MPU) based on a multicore CPU and a slave processing unit (SPU) implemented on an FPGA. Heterogeneous parallelization is used, with the MPU handling tasks like neighbor list building and data encoding/decoding, while the SPU performs the computationally intensive PES evaluations. High-speed data transfer between the MPU and SPU is achieved using techniques like full-duplex communication, PCIe, DMA, and a carefully designed pipeline. The paper details the algorithms for energy, force, and virial calculation within the NvN architecture, emphasizing the processing-in-memory (PIM) approach to minimize data shuttling. Specific details of the QNN, multiplication-less neural network, and modified activation function are presented, along with the hardware implementation using an FPGA.

The proposed NVNMD (Non von Neumann Molecular Dynamics) system demonstrates significant improvements in both accuracy and efficiency compared to existing methods. The root mean square errors (RMSE) for PES fitting across various molecules and bulk systems are well below the chemical accuracy threshold (1.0 kcal/mol), indicating high accuracy comparable to AIMD and MLMD. The calculation time efficiency of NVNMD is about two orders of magnitude better than MLMD, achieving CMD-level speed. Furthermore, the energy efficiency of NVNMD is 2-3 orders of magnitude higher than that of MLMD, largely due to the elimination of repeated data shuttling. The accuracy of atomic forces is also validated by comparing the results to those of AIMD and MLMD. MD simulations, including phase transition processes in GeTe and diffusion coefficient calculations in Li10GeP2S12, demonstrate the reliability and accuracy of the NVNMD methodology. The FPGA implementation, though based on a low-end device, serves as a proof-of-concept, suggesting even greater improvements are achievable with higher-end hardware like ASICs.

The authors acknowledge that the current implementation uses a low-end FPGA, limiting clock frequency and hardware resources. They anticipate significant further efficiency gains by migrating to an ASIC implementation, which should result in a substantial increase in both clock speed and available resources. The parallelization capabilities of the MLMD algorithm, and enhancements implemented in the pipeline, are expected to lead to order-of-magnitude improvements in efficiency. A comparison with Anton, another special-purpose MD computer, is provided, highlighting that NVNMD achieves AIMD-level accuracy unlike Anton, which uses classical force fields. The advantages of the FPGA-based implementation are also discussed, pointing towards its role as a valuable tool for research and testing before moving to ASIC-based mass production.

The study successfully demonstrates a novel MD methodology that achieves a significant advancement in both accuracy and efficiency. By utilizing a specially designed DeepMD algorithm adapted for a non von Neumann architecture, it surpasses the limitations of current AIMD and MLMD methods. The high accuracy and CMD-level speed are validated across various systems. While the FPGA implementation is considered an initial pilot study, it strongly suggests the potential for even greater performance enhancements with future ASIC implementations. Future work could focus on optimizing the algorithm for different hardware architectures and exploring new applications of this high-performance MD simulation approach.

The current implementation is based on a low-end FPGA, limiting the clock frequency and available hardware resources. The evaluation is primarily focused on a limited set of systems. A more comprehensive study involving a wider range of materials and system sizes is needed to fully assess the generalizability and scalability of the NVNMD method. Further optimization of the algorithm and hardware architecture is also possible. Finally, the cost and complexity of ASIC production compared to readily available CPUs/GPUs should be considered.