Deep Potential needed infrastructure for AI-for-science workloads whose compute demand could spike to hundreds of thousands of cores and whose data path had to support heterogeneous compute pools, hot small files, rapid loading, and quota management. Its Bohrium and Hermite platforms needed storage and scheduling patterns that could keep scientific AI jobs fast and economical even under extreme elasticity.
Reported impact
Alibaba Cloud helped Deep Potential turn AI-for-science infrastructure into a production system for industrial and biomedical R&D. The platform can handle hundreds of thousands of cores at peak, support cross-discipline data pooling, and shorten drug-screening and optimization cycles from years to months or even weeks while keeping storage and compute spending under tighter control.
Deep Potential and Alibaba Cloud designed a layered data infrastructure for AI-driven scientific computing and drug design. Near the compute tier, Fluid turns high-speed media inside the cluster into a logical cache layer so model files and data can be loaded faster and images can be pulled quickly enough to launch Pods in seconds. Above that, a general-purpose NAS layer provides shared read or write acceleration for many clusters and helps absorb hotspots caused by huge numbers of small files. For long-term retention, multi-tier OSS stores model files, checkpoints, and datasets economically while still supporting cross-region loading through Alibaba Cloud's global network. Those layers back Deep Potential's Bohrium research cloud and Hermite drug-design platform, allowing the company to pool data across heterogeneous compute, schedule large elastic jobs, and move from experiment-heavy workflows toward model-driven scientific discovery.
Scientific datasets, model files, and checkpoints are held in OSS, accelerated into NAS for multi-cluster sharing, and cached close to compute by Fluid when jobs are launched on ACK, ACS, and ECS. Those layers feed Bohrium and Hermite workflows so large-scale simulation, training, and drug-design tasks can load data quickly and schedule elastic compute resources efficiently.
Deep Potential layered the infrastructure into a near-compute cache with Fluid, a shared acceleration layer with NAS, and a long-term storage layer with OSS. (original: 数据缓存层... 使用阿里云 Fluid... 共享加速层... 采用阿里云通用型 NAS 方案... 长期存储层... 采用阿里云多级 OSS 对象存储方案。)
Deep Potential said its scientific-computing workloads can demand hundreds of thousands of cores and must provision compute and data resources in sub-minute or even second-level windows. (original: 科学计算任务对资源的需求具有极致弹性，峰值资源需求可达数十万核，需要在亚分钟级甚至秒级完成计算和数据资源的准备。)
The AI-for-science workflow needed heterogeneous compute, pooled data, low-latency handling of hot small files, and unified dataset management. (original: AI for Science 全链路涉及多种异构算力和数据池化的需求... 需要对热点小文件实现低延迟加速，并且在高频弹性任务中实现数据的极速加载和统一的数据集管理。)
With this platform, users can reduce drug screening and optimization work from years to months or even weeks. (original: 将原本需要数年才能完成的药物筛选和优化工作，用更大规模的算力协同调度缩短至几个月甚至几周。)
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Alibaba Cloud Sample Sep 17, 2026 Sinocare (三诺生物) Sinocare was rolling out multiple AI applications for chronic-disease management, including periodic assessment, anomaly analysis, glycemic prediction, diet imaging, and diet guidance, but it had to connect several… AI application launch speed improved 50% AI Gateway · Cloud Monitor 2.0 Alibaba Cloud Sample Archived · Mar 24, 2026 Cordis After receiving our products, distributors in China often request technical documentation, customs import declarations, and inspection & quarantine certificates. In the past, the China commercial team would forward… AnalyticDB · Model Studio · Qwen Alibaba Cloud Sample Archived · Oct 19, 2025 FathomX FathomX needed secure, compliant cloud compute and storage to train, validate, and deploy mammogram-analysis models at scale while controlling infrastructure costs and supporting regional expansion across Asian… 27.6% annually Elastic Compute Service · Elastic GPU Service 