推动研究前进的计算
- 云端
- 站在云端
- 按需
- 根据需要扩大规模
- 托管式
- 包括操作
- 已调优
- 特定于求解器的设置
核心能力
从设计到操作都在一个软件包中 - 因此研究人员可以专注于工作，而不是基础设施。
托管云 HPC
在云中设计和构建的专用集群。无需数据中心，无需内部管理团队。
每个解算器调整
针对 DFT、MD 和其他工作负载进行了调整的编译器、库和 MPI 设置 — 在相同的硬件上更快地获得结果。
突发容量
当内部 HPC 已满或队列很长时，请立即启动额外的节点，这样活动就不会停滞。
作业调度程序
基于标准 Slurm 的调度，具有优先级、配额和公平共享策略，用于团队级资源共享。
监控与运营
利用率、队列状态和作业历史记录的实时视图 - 事件响应由 VirtualLab 运营团队处理。
为你运行的内容付费
云计费意味着您按实际使用的时间付费，账面上没有闲置的硬件。
Getting results sooner from the same hardware
Cloud HPC performance depends far more on how the solver is built and how it communicates on those nodes than on how many nodes you switch on.
What tuning per solver actually means
The same source, built with a different compiler, linked against a different numerical library, running on a different MPI implementation, produces visibly different wall-clock times for an identical calculation. DFT codes are sensitive to FFT and linear algebra performance; MD codes to neighbour-list construction and inter-node latency. MAXFlops ships images with that combination already matched to the solvers your team actually runs, so nobody has to experiment with build flags.
How many nodes to attach
Parallel calculations do not run twice as fast on twice the nodes. Past some point communication cost overtakes the compute gain, and where that point sits depends on system size and solver. During onboarding we measure scaling on a representative calculation and set a sensible node count per job from it — over-parallelising costs more and returns results later.
Scheduling and sharing
A standard Slurm scheduler means existing HPC experience transfers unchanged. Priorities and quotas per team or project, plus a fair-share policy, stop one person's large campaign from blocking everyone else's queue. Utilisation, queue state and job history are visible in real time.
Data and cost
Results outlive the nodes that produced them, so storage and compute are designed with separate lifetimes. Resources stay up only for the hours they are needed, following cloud billing rather than leaving you holding idle hardware. Sizing starts from the solvers you run, concurrent users, data volume and budget, worked through together.
Why your existing workflow survives
Because the scheduler is standard Slurm, the submission scripts and habits from an in-house cluster mostly carry over unchanged. That was the point: learning a new command set or rewriting a pipeline should not be what stops a team from adding capacity. Where finished results are retrieved to, and how, is settled during onboarding — after which the aim is that a researcher cannot tell whether a job ran locally or in the cloud.
如何交付
我们倾听您的工作负载，为其设计一个集群，然后启动它。从那里开始，你就可以使用它——这就是整个流程。
-
01
定义工作负载
您主要运行的解决方案、并发用户、数据大小和预算 - 我们一起解决这些问题。
-
02
设计与站立
实例、网络、存储和软件堆栈在云中设计，针对您的求解器进行调整，并上线。
-
03
运行和操作
研究人员只需提交工作即可。监控、事件响应和横向扩展由 VirtualLab 处理。
团队在哪里使用它
大规模物资宣传活动
对许多候选结构和数千个作业数据生成管道进行 DFT/MD 筛选。
AI训练数据生成
材料 ML 的计算数据集 — 必须在截止日期内完成的工作负载。
日常 DFT 和 MD
将因内部HPC排队而停滞的日常计算按需转移到云端。
短期合作和项目
在项目期间启动专用资源，然后拆除——无采购周期。
MAXFlops — frequently asked questions
What is MAXFlops?
A managed service that designs, builds and operates a dedicated HPC cluster for your team on the cloud. You get a large-scale compute environment without procurement, a data centre, or a dedicated administrator.
We already have in-house HPC. Why would we need this?
As headroom. When in-house resources run short or queues grow long, extra nodes come online immediately so a calculation campaign does not stall — expanding for the period you need instead of permanently.
How are jobs submitted and resources shared?
Through a standard Slurm scheduler, with per-team priority, quota and fair-share policies. Utilisation, queue state and job history are visible in real time.
What does "tuned per solver" mean?
Compiler options, numerical libraries and inter-node communication are optimised for the solvers you actually run — DFT, MD and others. The same resources return results sooner, and researchers never have to tune the environment themselves.
How is the cost determined?
Cloud billing for the hours the resources are actually up. We size the cluster after going through the solvers you run, concurrent users, data volume and budget together.