Alibaba Cloud's You Liang Discusses Hyper-Convergence of HPC and AI: How E-HPC Unifies Scheduling of HPC, AI, and Agents

If we think of scientific research computing as an assembly line, then for the past few decades the rhythm of that line has been controlled by people: researchers find tools, write scripts, submit jobs, watch the queue, debug and rerun, and finally summarize the results grid by grid. The computing power itself is powerful, but the way tasks are fed to it has remained stuck in the hand-crafted era.

You Liang, head of Alibaba Cloud's High-Performance Computing & Performance Acceleration, told us that in the current research workflow, more than 80% of the time in stages such as data preparation, environment configuration, task orchestration, and result analysis still depends on manual handoffs, with less than 20% of the time actually spent on computational execution.

The convergence of supercomputing and intelligent computing is now an inevitable trend

In recent years, several Nobel Prize-winning achievements have been directly related to AI. AI is no longer merely an auxiliary tool for research; it is gradually entering every stage of observation, simulation, conjecture, and experimentation, becoming a key component of the research workflow. At the same time, the high-precision scientific computing of traditional HPC and the training and inference of AI models are now appearing in parallel within the same research project: the same study cannot do without either high-precision physical simulation or large-scale model training. The "China AI Computing Power Development Assessment Report" jointly released by IDC and Inspur shows that in 2024 China's intelligent computing power reached 725.3 EFLOPS, up 74.1% year on year, far higher than the 20.6% growth of general-purpose computing power; the report predicts that by 2028 China's intelligent computing power will reach 2,781.9 EFLOPS, maintaining long-term rapid growth. Another intelligent computing service market tracking report released by IDC also shows that in the first half of 2024, China's intelligent computing service market grew by nearly 80% year on year, with generative AI infrastructure services growing by more than 200%.

In You Liang's view, as multiple types of computing rush onto the same computing platform at the same time, it is very difficult for traditional HPC architectures to achieve optimal performance and lowest cost for all of them simultaneously. What is even more troublesome is scheduling: the scheduling logic of traditional HPC is job-based. Slurm or PBS knows how many cores and how much memory a task needs and how long it is expected to run, but it does not know what scientific problem the task is meant to solve. When a bioinformatics researcher submits a variant detection job, what the scheduler sees is a computing job; it cannot see which link in the complete chain from FASTQ to candidate mutations the job occupies, whether to switch tools or adjust parameters after a failure, or what should be run next once the results come out—all of these judgments are left to humans.

Now, however, the system must simultaneously schedule HPC jobs, AI training and inference, agent tasks, and sandbox tasks. The characteristics of these workloads differ enormously: high-precision simulation consumes double-precision floating-point performance and memory bandwidth, AI training consumes interconnect bandwidth and VRAM capacity, and agents' tool calls and sandbox tasks consume lightweight, fast start-stop, and elastic resources. A traditional HPC cluster architecture designed for a single type of workload cannot make every type of computing run optimally, and at this point the traditional single-type scheduling approach clearly can no longer hold up.

Today, the research workflow is moving from humans driving tools to humans setting goals and agents executing autonomously, and this new workflow's requirements for computing power are: it must run HPC, run AI, and run agents, and it must have unified scheduling and unified metering. Whoever can bring these three workloads with different appetites to one table will get the ticket to the next generation of research infrastructure, and this is also why the convergence of supercomputing and intelligent computing has been pushed to the forefront.

Where is the difficulty in the convergence of supercomputing and intelligent computing?

The convergence of supercomputing and intelligent computing looks like simply connecting supercomputing servers and intelligent computing servers to work together, but actually doing it is not that simple. The most intuitive difficulties lie in precision and data flow. Weather forecasting, fluid simulation, and molecular dynamics usually require FP64 double precision to suppress accumulated errors, while large-model training and inference make extensive use of FP16, BF16, or even FP8, and agents' sandboxes and tool calls involve a large amount of fragmented, short-lived lightweight computing. Traditional computing systems separate high-precision computing from low-precision computing, making it difficult to meet the needs of mixed-precision computing, AI4S, and other composite business requirements. Precision is not a matter of "choosing a gear"; rather, within the same research chain, it is necessary to switch seamlessly between different gears.

You Liang shared Alibaba Cloud's solution with us, saying that it can be broken down into three layers.

First layer: let different types of computing power each play to its strengths while connecting into one computing network. You Liang explained that Alibaba Cloud's approach is to provide the most matching instance types for different workloads. For example, for traditional HPC it provides HPC-optimized instances; for AI training and inference it provides corresponding AI-accelerated instances respectively; and for agents it provides general-purpose computing and container instances, among others. The key is that these heterogeneous instances are linked into one "computing network" through a unified virtual network on the cloud, and with the CIPU 2.0 architecture they achieve resource scheduling with no virtualization loss, and through a unified open OpenAPI all computing power can be uniformly created, scheduled, and orchestrated.

At the CPU level, Alibaba Cloud has built hpc9a, which is equipped with fifth-generation AMD EPYC (Turin-C) processors, with a maximum boost frequency of 5 GHz, 128 physical cores, and memory of up to 3 TB. Relying on the CIPU 2.0 architecture to eliminate virtualization overhead and paired with a 200G low-latency eRDMA network, it delivers about a 30% performance improvement over the previous generation.

You Liang introduced that hpc9a is specifically designed for scenarios such as CAE simulation and electronic design automation that are extremely sensitive to compute performance, memory bandwidth, and network latency. On the AI side, T-Head's self-developed chips delivered results: the M890P improves performance to 3x that of the previous generation, and the new-generation V900 further improves performance to 3x on top of that. Combined with the supernode architecture and elastic computing power, E-HPC can provide AI acceleration capabilities for scientific research scenarios ranging from a single card to large-scale clusters.

In intelligent computing, Alibaba Cloud has also achieved results on the path of self-developed chips: its M890P improves performance to 3x that of the previous generation, and the new-generation V900 further improves performance to 3x on top of that. Combined with the supernode architecture and elastic computing power, E-HPC can provide AI acceleration capabilities for scientific research scenarios ranging from a single card to large-scale clusters.

Balancing performance and cost relies on two approaches. First, multiple generations of instances coexist, and customers choose the instance specification with the best cost-effectiveness for their workloads; second, the elasticity of the cloud means that customers may need thousands of instances at business peaks and perhaps only dozens at troughs. The traditional model must reserve resources according to peak demand, while on the cloud resources can be created and destroyed on demand, bringing both performance and cost close to optimal. You Liang shared a case with us: Alibaba Cloud worked closely with a leading domestic passenger car company and, through HPC-optimized instances, elastic eRDMA low-latency networking, and CPFS parallel storage optimization, helped the customer save about 25% of the computing cycle end to end, greatly improving work efficiency.

The second layer is to make data travel faster than computing power. No matter how strong the computing power is, if data cannot be fed in, it is just idling. You Liang introduced that eRDMA is responsible for making communication between nodes increasingly fast, while CPFS is responsible for enabling workloads such as HPC computing, distributed AI training, and inference to obtain data in the shortest time, keeping computation saturated and minimizing waste of expensive computing power. Storage-side optimization includes hot-cold tiering, placing frequently used data in the cache layer so it does not need to be pulled from remote storage when used, as well as affinity scheduling for compute nodes that are network-near and storage-near.

The third layer is to let agents complete the execution process on people's behalf. Around trends such as the continuous integration of supercomputing and intelligent computing, the rising role of AI in scientific research, and manual processes becoming bottlenecks, Alibaba Cloud launched the E-HPC agent platform, whose architecture is divided into four layers from top to bottom: the top layer consists of agents for industries such as embodied intelligence, life sciences, CAE, EDA, quantitative finance, and rendering; the industry access layer distills industry knowledge and expert methods into Skills for agents to call, while also providing a command line, open APIs, and SDKs; in the middle is a unified Agent Runtime, responsible for understanding research goals, planning tasks, actually executing with skills, and re-planning based on logs and result feedback; below that is the E-HPC execution layer, which uniformly handles job orchestration, heterogeneous resource scheduling, environment and data preparation, and state log management, and schedules CPU, GPU, eRDMA network, image software, and security governance resources.

By building these three layers of architecture, Alibaba Cloud has already achieved results in supercomputing-intelligent computing integration. In the embodied intelligence scenario, RoboAgent connects the complete chain from device-side data collection, data processing and annotation, model training and inference, remote simulation, to deployment feedback, reducing data preparation costs in a one-stop manner; a leading company in the embodied field used this solution to achieve a rapid closed loop from real-machine collection and upload, multimodal annotation processing, to training of multiple algorithms such as QwenVL.

In the life sciences scenario, BioAgent automatically orchestrates WDL workflows according to research goals, autonomously identifies data formats such as FASTQ and PDB and recommends molecular simulation parameters, calls tools such as GATK and GROMACS, and intelligently matches CPU/GPU resources, making bioinformatics and drug research processes reusable, interpretable, and continuously iterable. The most representative example is "Huntianling" from Gewu Zhiyan under the Shanghai Academy of AI for Science. This intelligent R&D platform for materials science, which orchestrates more than 140 scientific research skills, works with Alibaba Cloud E-HPC's elastic computing power foundation to take molecular screening tasks from data retrieval, computing power invocation, molecular docking, and FEP calculation to automated synthesis. These steps, previously linked manually, are now all completed autonomously by agents.

In the CAE scenario, engineers only need to focus on design judgment, while preprocessing, solving, diagnosis, post-processing, and parameter exploration are autonomously connected by the Agent.

Achieving these goals is not something that can be done overnight. You Liang also admitted that the biggest challenge is that different industries have different requirements for computing power, precision, and implementation methods. It is necessary to distill the expert experience and industry knowledge of various industries into skills, while also distilling E-HPC's expertise in scheduling, orchestration, task identification, and resource matching into platform skills, and then undertake both through a unified Agent Runtime and solve difficulties such as the collaboration between industry skills and E-HPC skills, the launch of massive computing power, and the optimal matching of industry workloads with computing workloads. His summary is: industry Agents can call Alibaba Cloud's HPC services in a very stable, elastic, observable, and governable manner, and behind this, every layer needs repeated refinement.

More noteworthy is that E-HPC is no longer just a general-purpose computing platform, but has formed dedicated solutions for three key industries: chip design, AI4S and life sciences, and embodied intelligence. For chip design, E-HPC hybrid cloud can seamlessly connect to enterprises' existing LSF, Slurm, and PBS scheduling environments, with unified scheduling on and off the cloud and automatic scaling by queue, so engineers can supplement computing power at any time during simulation and verification peaks without changing their usage habits; for large-memory jobs such as STA and PR, large-memory HPC instances can directly handle them, helping chip teams safeguard tape-out cycles and optimize total cost of ownership. For AI4S and life sciences, E-HPC pre-integrates mainstream research software such as AlphaFold3, OpenFold3, RFDiffusion, GROMACS, and GATK, ready to use out of the box, and customers' own research Agents can also directly call elastic computing power through CLI, API, or Skill, freeing drug discovery and genetic analysis from cumbersome environment setup. For embodied intelligence, HPDA Robotics strings collection, management, annotation, training, simulation, and inference into a complete data closed loop, with a built-in APAK acceleration engine that can improve the cost-effectiveness of multimodal data processing by 2.1 to 5 times, significantly shortening data preparation and model iteration cycles. The three dedicated solutions share the same elastic computing network and the same set of scheduling, storage, and security capabilities, and can be used out of the box as well as smoothly integrated into customers' existing environments.

Agents will become the operating system for scientific research

Through the three-layer architecture built by Alibaba Cloud, the current development path of super-intelligence integration can be seen. On the technology side, computing power will continue to grow, and transmission channels will continue to lengthen. You Liang revealed that in the Scale Up direction, super nodes on Alibaba Cloud consist of dozens to hundreds of AI chips, and future super nodes will support larger scales; in the Scale Out direction, cross-node and cross-availability-zone interconnection is achieved through an elastic RDMA network.

A super node can essentially be viewed as a giant AI computing node, and the continuous raising of its scale ceiling will directly determine the efficiency of large-scale training and simulation. At the same time, the difference between eRDMA and traditional RDMA is that it can carry broader interconnection, extending from inside super nodes to cross-availability-zone scenarios, which is precisely the physical foundation of the "computing network" imagination.

On the application side, Agents will change from an optional item to an industry standard. The product logic of Alibaba Cloud E-HPC has moved from single-point Agents to systematization: Agents in various industries are responsible for professional execution in their respective domains, while WorkAgent converges the Agents scattered across industries into a unified goal-driven execution loop. Researchers only need to clearly state "what to compute, how accurate it needs to be, how long until results are available, and how much it costs," and WorkAgent is responsible for translating this into computing tasks and acceptance criteria, automatically decomposing processes, introducing expert experience, matching computing network and storage resources, and finally parsing logs, convergence curves, and key metrics, locating failures, and giving improvement directions. "People only need to give goals and make judgments, and repetitive workflows are left to Agents to complete autonomously. The significance of this step is that enterprises do not need to repeatedly build industry Agents and computing foundations, but can access them like water and electricity," said You Liang.

Alibaba Cloud is one of the most representative companies in the field of super-intelligence integration, but it is not the only one. In the traditional supercomputing camp, national supercomputing centers and hardware vendors are advancing the deep integration of HPC and AI methods along their respective platforms; AI vendors focus on intelligent computing clusters and ten-thousand-card training platforms; and telecom operators are building intelligent computing centers on a large scale. IDC data shows that the intelligent computing integration services market is expanding at a growth rate of more than 150%, and the market landscape is far from settled.

It is worth noting that although each company's approach is different, the direction is highly consistent: let HPC and AI work collaboratively on the same platform, and let Agents become the default operating system for scientific research.