It takes several years for new materials to go from laboratory R&D to actual commercial application, and the core bottleneck lies in the repeated rounds of trial testing: Xipeisen intends to use atomic-level dedicated computing power to change the first step of R&D from "going into the laboratory" to "performing relevant calculations first".

How long does it take for a new material to go from the laboratory to practical application? For a very long period of time, the answer to this question was determined by the number of experiments — rounds of formulation adjustment, sample preparation, and performance testing, and it is not uncommon for the whole process to take several years.

The cost of the trial-and-error method is far more than just time

The cost of the empirical trial-and-error method is not only reflected in the experimental expenses. It can hardly answer the "why" question — which micro factor exactly contributes to the better performance of a certain formulation often cannot be quantitatively explained. A successful formulation often can only stay at the success itself, cannot be precipitated into reusable rules, and is difficult to be systematically optimized.

This means that almost every new material R&D process starts from scratch. The experience accumulated by the research group is difficult to transfer between different material systems. A formulation idea that works for alloy materials may not be applicable to ceramic materials; a process parameter verified in a polymer system may be completely invalid when applied to an inorganic non-metallic system. The efficiency bottleneck of the trial-and-error method is essentially the bottleneck of the knowledge accumulation mode.

Where have computational methods advanced to

The development of computational materials science provides another solution to the above problems. Density Functional Theory (DFT) and Molecular Dynamics (MD) are two widely applied representative computational routes: the former is used to solve electronic structures and obtain the ground-state properties of materials from the quantum mechanics level with high precision, but the computational volume rises sharply with the scale of the system; the latter presents the dynamic behavior of materials at finite temperature by integrating the atomic motion equations, with low single-step computational cost and larger processable system, but its precision depends on the quality of the potential function, and the physical time that can be simulated is limited by the integration step size.

In recent years, the application of Machine Learning Molecular Dynamics (MLMD) has found a new balance between computational efficiency and simulation precision. MLMD replaces the traditional empirical potential function with the potential function trained by machine learning. While maintaining the precision close to DFT, it significantly expands the scale and time scale of the simulatable system, which is one of the most concerned progresses in computational materials science in recent years.

However, after the computational methods continue to mature, the factors restricting them gradually shift from algorithms to hardware. No matter how good the algorithm is, it needs to run on computing power to produce results.

Two walls: the bottleneck of general-purpose computing power

Atomic-level simulation is a rather "demanding" computing workload: it features large system scale, numerous iteration steps, and intensive data throughput. Taking molecular dynamics simulation as an example, for a system at the order of millions of atoms, the interaction forces between all atoms need to be calculated at each time step; if the first-principles molecular dynamics based on Density Functional Theory is adopted, these forces need to be derived from the solution of electronic structures at each time step, and the computational volume rises sharply with the number of atoms.

However, general-purpose CPUs/GPUs are designed for general computing scenarios. Under the general architecture, "storage" and "computation" are naturally separated: data is stored in the memory, and needs to be moved to the arithmetic unit before participating in the calculation. The movement itself does not produce calculation results, but it does consume real computing power and energy consumption — when the data throughput speed cannot keep up with the computing demand, the memory wall is formed; when the movement energy consumption rises with the system scale, the power consumption wall is formed.

The superposition of the two walls limits both the speed and energy efficiency of material computation, making it difficult to implement large-scale system atomic simulation on general-purpose computing power. Improving hardware specifications at a single point — using a larger video memory or stacking more GPUs — can alleviate the symptoms, but cannot solve the problem fundamentally, because the bottleneck does not lie in the quantity of hardware, but in the design logic of the architecture.

Redesign computing power for specific scenarios

The path chosen by Xinpeisen is to redesign the computing power from the architecture level. Xinpeisen APU (Atomistic Processing Unit, a dedicated computing power unit for atomic-scale material computation) adopts an original "non-Von Neumann" dedicated chip architecture, which allows data to participate in the operation right next to the storage location — this is the core idea of this architecture at the bottom layer, and also the key to reducing the overhead of data movement.

Dedication does not mean exclusivity. The Hexi server with Xinpeisen APU as the core adopts a super-heterogeneous computing architecture (CPU + GPU + APU collaboration), which can support AI model training, AI model inference, DFT calculation and MD calculation. This means that when processing dedicated workloads such as atomic simulation, APU undertakes the main computing tasks; when general computing is required, CPU/GPU works as usual. The two types of computing power are collaboratively scheduled in the same server, rather than replacing each other.

According to the person in charge of Xinpeisen, in atomic-level scientific computing scenarios such as Machine Learning Molecular Dynamics (MLMD) and Density Functional Theory (DFT), compared with general-purpose CPU/GPU (comparison benchmark: 1 A800 GPU with 80GB video memory), APU has significantly improved computing speed and significantly reduced energy consumption. Taking the high-entropy ceramic impact simulation of about 1.1 million atoms as an example, the impact simulation task of this scale can be completed with only one computing node in the Hexi cluster; in this scenario, with the help of Xinpeisen's machine learning potential function training method, the potential function training cost is only one third of that of the traditional solution.

Beyond computing power: tools and services

Computing power is only the first threshold. For the paradigm shift to truly happen, two more things are needed: first, tools that allow researchers to no longer spend a lot of time on manual modeling, repeated parameter adjustment, and troubleshooting errors; second, services that make the calculation results reliable and traceable.

Xinpeisen's supporting facilities in these two directions are the intelligent agent (Xinpeisen · MaterPlato) and computing service (Xinpeisen · SwiftMater) respectively: the former undertakes automatic modeling, parameter optimization, error pre-judgment and auxiliary post-processing, lowering the threshold for use; the latter is directly connected to users by a full-time technical team of master and doctor degree holders. Together with its super terminal (Xinpeisen · MaterHaven), the two form a "ternary closed loop".

The change of paradigm

The paradigm shift of material R&D from empirical trial and error to scientific computing will not be completed with a single product release. It is more like a slow variable: when the algorithm matures one step, the computing power advances one step, and the tool becomes easier to use one step, the work habits of researchers may change a little accordingly. But a slow variable does not mean standing still — every reduction in the threshold shortens this distance.

The indicators for measuring such a shift are actually very simple: when researchers encounter a material problem, their first reaction is to go to the laboratory to do a round of experiments, or to calculate on the computer first. Where this first reaction comes from depends on whether three things are in place in the daily process: whether the computing power is available on demand, whether the tool requires no additional learning cost, and whether the calculation results are reliable and traceable. These three things exactly correspond to the three links of computing power, tools and services; the paradigm shift of material R&D is ultimately reflected in the change of this first reaction.