Industry Cases

Customer Background

The existing container platforms managed only traditional workloads, GPU resources could not be scheduled across clusters, and monitoring, storage, and logging infrastructure was being duplicated.

Customer Pain Points

• 01 GPUs across multiple clusters cannot be scheduled uniformly
• 02 Domestic GPUs need finer slicing and scheduling capability
• 03 Duplicated monitoring, storage, and logging infrastructure

Solution

Adopting a control-plane / GPU-compute separation architecture, using Rise CAMP to integrate multiple domestic and NVIDIA clusters with the existing container platforms, reusing LDAP, image registry, storage, and monitoring, and enabling dynamic slicing on domestic GPUs.

Customer Background

As large models move into core scenarios, traditional architectures struggle to support multi-team concurrent development and fast iteration, and the compute platform must deeply integrate with the domestic-compatible software stack.

Customer Pain Points

• 01 Multi-team concurrency and fast iteration place greater demands on compute scheduling
• 02 Deep compatibility with the domestic-compatible ecosystem is required
• 03 End-to-end training and inference lifecycle management is needed
• 04 Small-model fine-tuning capability for specific business scenarios is needed

Solution

Building a full-stack domestic-compatible heterogeneous compute base on Rise ModelX, leveraging Rise VAST fine-grained slicing and dynamic scheduling to deliver one-click standardized workspaces, with a built-in model marketplace and inference engine plus support for business-scenario fine-tuning.

Customer Background

Financial AI workloads are driving a surge in compute demand, creating an urgent need for an autonomous and controllable compute base. Multi-team parallel workloads on heterogeneous NVIDIA and domestic GPUs have formed compute silos, and whole-card passthrough leaves utilization low.

Customer Pain Points

• 01 Resources across the two-region three-data-center topology cannot be managed uniformly
• 02 Heterogeneous GPU scheduling barriers have formed compute silos
• 03 Whole-card allocation cannot match fine-grained slicing needs
• 04 Business workloads require zero-downtime smooth migration

Solution

Using Rise CAMP as the heterogeneous compute base to unify management across the three data centers, hide chip-level differences, support fine-grained slicing down to 1% compute and 1MiB VRAM, and deliver zero-downtime smooth migration.

Customer Background

With multi-team parallel workloads and no GPU virtualization in place, heterogeneous GPU utilization stayed low, and the bank lacked visualized management and unified development standards.

Customer Pain Points

• 01 Compute demand keeps growing while GPU utilization has significant room to improve
• 02 Already-allocated resources lack visualized oversight
• 03 Compatibility between GPU pooling and existing application environments needs to be strengthened
• 04 Development frameworks and tooling require unified standards

Solution

Using Rise VAST to consolidate existing compute, delivering unified GPU management, overcommit, and elastic scaling, establishing an operations framework, and standardizing development tooling integration.

Customer Background

Each department ran its own dedicated hardware as siloed stacks, with no unified authentication or auto-scaling in place, leaving utilization very low.

Customer Pain Points

• 01 Resources are scattered across departments and need unified management with heterogeneous scheduling
• 02 A unified permissions and monitoring framework is needed to support standardized operations
• 03 Model deployment still relies on manual adjustments and needs automated elastic scaling

Solution

Using Rise CAMP on Kubernetes to unify management of diverse domestic heterogeneous resources, providing container/task/tenant/quota/metering, GPU slicing and overcommit, unified scheduling, and full-stack observability.

Customer Background

The supercomputing center needed precise compute scheduling for scenarios like exam monitoring. The original GPU allocation was by physical server, leaving single-host utilization low, significant idle capacity outside exam periods, and heavy reliance on manual O&M.

Customer Pain Points

• 01 Whole-node allocation cannot fully unlock the value of high-end cards
• 02 Large amounts of compute sit idle outside exam periods
• 03 Resource scheduling relies on manual effort, with considerable O&M overhead

Solution

Deploying Rise CAMP scheduling platform to connect exam video streams and NFS storage, pooling GPUs for task-level allocation, and automatically redirecting idle compute to research workloads outside exam periods.

Customer Background

Under a pure-domestic compute environment, full-scale large models must run privately on-premises, and multi-department heterogeneous domestic hardware can easily lead to resource fragmentation.

Customer Pain Points

• 01 Pure-domestic compute must support full-scale large models
• 02 Multi-vendor heterogeneous hardware leads to resource fragmentation
• 03 A unified model service and fine-tuning environment is needed
• 04 Cross-department unified metering and attribution is needed

Solution

Rise CAMP seamlessly unifies multi-vendor domestic heterogeneous compute into a single pool, supports fine-grained slicing of compute and VRAM, and exposes a unified API gateway that hides backend complexity and provides standard authentication.

Customer Background

The group's MaaS platform could not elastically provision compute on demand, and multi-team parallel workloads on heterogeneous NVIDIA and domestic GPUs with whole-card passthrough had formed compute silos.

Customer Pain Points

• 01 The telecom MaaS platform lacks unified compute scheduling
• 02 Heterogeneous GPU resources urgently need unified management
• 03 A metering, billing, and resource monitoring framework is still to be built

Solution

Using Rise CAMP to build a unified heterogeneous compute management platform that connects to the telecom MaaS, with a base supporting multi-vendor accelerator integration.

Customer Background

Baidu MaaS and Alibaba MaaS are deployed on different clusters with siloed resource pools that cannot be shared; management granularity stops at the physical card level, making a unified compute platform difficult to build.

Customer Pain Points

• 01 Multi-vendor MaaS resource pools are isolated and hard to share
• 02 Management granularity stays at the physical card level and adapts poorly to small-model scenarios
• 03 A lack of dynamic scheduling leaves room to grow overall utilization

Solution

Using Rise VAST to unify both MaaS platforms, applying heterogeneous pooling and virtualization to connect cross-cluster resource scheduling and deliver a fine-grained, elastic, migratable compute base.

Customer Background

Large-model training nodes are tightly coupled, and traditional scheduling easily leads to global stalls and idle compute; long-duration training is prone to failures, and the trade-off between heavy road-test data I/O and storage TCO is sharp.

Customer Pain Points

• 01 Distributed scheduling deadlocks leave compute idle
• 02 Long-duration training demands high fault tolerance; failures can abort the entire job
• 03 Tension between massive-data I/O throughput and storage TCO

Solution

Using Rise CAMP on the GPU cluster with the Volcano scheduling framework and automatic fault tolerance, backed by all-flash and hybrid-flash tiered storage for hot and cold data, improving training efficiency and cost control.

Customer Background

Building an aviation-industry large-model incubation base required unified management of multi-vendor GPUs and a visual agent platform integrating various open-source models.

Customer Pain Points

• 01 NVIDIA and domestic Ascend compute differ significantly at the underlying layer, making heterogeneous cards hard to manage uniformly
• 02 Small-model development monopolizes whole cards, leaving compute underutilized; finer allocation is needed
• 03 The existing platform struggled to reliably support high-concurrency access across the campus

Solution

Deploying Rise CAMP and the agent platform, using domestic GPUs for dynamic proportional slicing, and applying a shared/dedicated dual-pool model to unify open-source model management across research, teaching, and digital human scenarios.

Customer Background

Clinical department data must be strictly isolated, multi-vendor workloads sharing GPUs make whole-card allocation impractical, and lengthy AI environment setup slows delivery.

Customer Pain Points

• 01 Whole-card allocation does not match the fine-grained compute needs of small models
• 02 Strong data isolation requirements across departments
• 03 Whole-card allocation cannot serve multiple vendors sharing the same infrastructure
• 04 Lengthy AI environment setup impacts development delivery pace

Solution

Deploying Rise CAMP on the GPU cluster to provide GPU pooling with fine-grained slicing and multi-tenant isolation, and integrating development tooling that accelerates AI delivery.