Frequently Asked Questions

Follow the "bottom-up, additive" principle: Rise VAST handles GPU virtualization and heterogeneous compute pooling — the foundation everything else builds on. Rise CAMP sits on VAST and adds intelligent task scheduling, cluster and multi-tenant management, quota management, compute resource monitoring, and fault diagnosis. Rise ModelX delivers a unified training-and-inference model serving platform with an AI Gateway. Rise MAX packages all three layers plus hardware and pre-loaded models as a turnkey appliance.

Quick decision logic:

• Greenfield, want to skip integration risk → Rise MAX appliance
• Existing GPU cluster, low utilization → Start with VAST to drive utilization up
• Existing GPU cluster, scheduling chaos → CAMP for fine-grained scheduling
• Need an AI development environment → CAMP for AI dev services
• Model inference for internal and external consumers → ModelX standalone
• Large production cluster, full MLOps → VAST + CAMP + ModelX stacked

Each layer is independently purchasable and deployable. The typical evolution path:

1. Phase 1 — deploy VAST to push existing cluster utilization from 30% to 70%+, prove the cost savings
2. Phase 2 — add CAMP for multi-team scheduling governance, boosting task scheduling efficiency
3. Phase 3 — layer in ModelX to unify internal model serving

Reverse paths also work — start with ModelX to consolidate model serving, then push VAST/CAMP underneath later. Every layer is Kubernetes-native and hot-swappable; no cluster rebuild required.

For most customers, starting with VAST gives the highest ROI. Here's why: the stock K8s Device Plugin only does whole-GPU allocation. In real clusters, inference services, Notebooks, and small training jobs can't saturate a full card — industry-average GPU utilization stays around 30%. VAST's vGPU virtualization lets multiple Pods share a single card, typically pushing utilization to 70%+. Effectively, your existing GPU count "doubles." This can significantly improve the return on hardware investment.

Further reading: GPU Virtualization Technology Guide · GPU Pooling for Accelerated Training

No. Rise ModelX deploys standalone on any standard Kubernetes + GPU cluster, delivering model registry, inference engines, AI Gateway, and token governance out of the box. But VAST makes it work better — VAST's vGPU partitioning lets a single physical card host multiple inference services concurrently, and combined with ModelX's smart routing, overall serving ROI improves significantly.

Three delivery models:

1. Software subscription — customer provides hardware, we handle install, tuning, upgrades, and support. Best for customers with existing GPU clusters or strong hardware preferences.
2. Rise MAX turnkey appliance — integrated hardware + software + pre-loaded models, operational in under an hour. Best for customers who want fast time-to-value.
3. End-to-end private cloud deployment — for large customers, we handle data center planning, hardware procurement, network buildout, and software deployment.

No public cloud SaaS offering currently — all products are designed for private deployment to keep data inside the customer perimeter.

Typical PoC runs 2-4 weeks covering environment prep, functional validation, performance testing, and report delivery — with one-click install, core platform components deploy in minutes.

The platform uses a "control-plane and compute-plane separation" architecture: the control plane runs on Kubernetes-native components. Minimum footprint: a single node is enough to kick off a PoC — the control plane and GPU compute plane can co-locate on the same machine, bare-metal or virtual machine, as long as GPU resources are attached. PoCs can run on a subset of an existing cluster or a standalone test environment.

Software deployment requires: standard Kubernetes cluster (1.20+), GPU nodes, network and storage, and ops access (SSH/bastion). Appliance deployment requires: standard rack space, power (220V / 32A dual feed), and network uplink. Day-to-day ops have a low bar — every component has a web console and full observability. Routine operations (creating tenants, adjusting quotas, viewing monitoring) don't require deep Kubernetes expertise. Deep tuning and incident response are handled remotely by certified RiseUnion engineers. Operations training and certification programs are available.

Recommended production topology: 3-node HA control plane, compute plane scales independently and supports clusters of up to 500 GPU physical servers. Larger environments should add additional clusters. Storage platforms, container registries, and related infrastructure should be sized based on workload and scale.

Typical delivery takes 2-4 weeks, covering architecture design, environment preparation and validation, production deployment, functional verification, performance testing, report delivery, and on-site training. Ongoing 7×24 support from certified RiseUnion engineers follows.

Yes. Rise CAMP provides federated cluster management, unifying multi-region, multi-data-center, and multi-vendor Kubernetes clusters under a single control plane with cross-cluster scheduling, resource pooling, and unified observability. Common patterns:

1. Multiple GPU clusters across geographically distributed data centers that need centralized management and unified compute service delivery
2. Train in DC, infer at edge — combined with Rise Edge for centralized training + edge inference + OTA model distribution
3. Hybrid cloud burst — overflow to public cloud when local capacity is saturated, transparent to applications

All components support rolling upgrades, canary releases, and one-click rollback — workloads stay running throughout. Two to three major releases per year (with new features), plus continuous security patches, all bundled into annual maintenance. Pre-upgrade compatibility validation ensures any historical version can upgrade smoothly to the latest. Upgrades are performed remotely or on-site by RiseUnion engineers — no breaking changes that require cluster rebuilds. Self-service upgrades are also supported with detailed runbooks and rollback procedures for every release.

Yes. With Rise ModelX, we provide inference deployment best practices for mainstream open-source models, including recommended engine selection (vLLM / TGI / TensorRT-LLM, etc.), VRAM and compute sizing guidance, concurrency and batching tuning, and AI Gateway load-balancing and autoscaling configurations. For popular model families like DeepSeek, Qwen, and LLaMA, we have production-validated deployment templates ready to use.

For non-standard models (custom-trained, privately fine-tuned, or uncommon-framework models), we offer three tiers of support:

1. Self-service onboarding — ModelX accepts any inference backend that exposes an OpenAI-compatible API; customers can integrate following our documentation.
2. Remote assistance — RiseUnion engineers help with model adaptation, engine tuning, and performance baselining.
3. Deep customization — for custom operators, bespoke inference pipelines, or other complex scenarios, both parties jointly assess the technical approach; additional fees may apply, with specifics agreed during PoC or implementation.

Per-GPU license purchase or subscription, tiered by GPU class. Contact sales for pricing (contact@riseunion.io).

Three-tier first-party support:

1. L1 ticket response — 30-minute response for critical issues, 4-hour for normal issues
2. L2 remote engagement — root cause within 4 hours
3. L3 on-site support — 24-hour arrival for major incidents

On-site resident engineering is available for key accounts.

Default SLA covers business hours 9×5; 24×7 year-round support is available as an upgrade. All support is delivered by RiseUnion engineers directly — no outsourcing, no subcontracting. Quarterly health checks, annual optimization reviews, and proactive upgrades are included.

Free PoC, 2-4 weeks, including solution design, environment deployment, functional validation, performance testing, and a final report. RiseUnion engineers provide on-site or remote support throughout. All test data and findings belong to the customer. For longer trials (1-3 months), we sign a trial agreement and provide a time-limited license. PoC pass rate exceeds 90% — most customers validate the GPU utilization gains and cost reductions during the trial itself.

Three-tier training program:

1. User training — for developers and ML engineers, 1 day, covering job submission, model serving, and monitoring.
2. Operations training — for IT operations, 2-3 days, covering daily operations, quota management, and troubleshooting.
3. Paid advanced training — for architects and power users, 3-5 days, covering scheduling policy tuning, performance optimization, and customization.

Training is delivered on-site or remotely; certifications are issued upon completion.

We have a comprehensive partner program. Contact contact@riseunion.io for details.

Products are running in production at scale across finance, government, transportation, education, healthcare, and energy. Publicly disclosed reference customers include BYD (autonomous driving model training), Huatai Futures (financial AI inference), National Supercomputing Center in Jinan (public compute services), Second Affiliated Hospital of Chongqing Medical University (medical AI), and CEPREI Laboratory (research MaaS platform). Detailed cases and industry solutions are on the Industry Cases page.

All public references are listed on the Industry Cases page, organized by vertical, with technical solution, deployment scale, and outcomes. For prospective customers, with the reference customer's authorization, we arrange phone or on-site discussions with same-industry customers to help evaluate implementation risk and benefit. Contact sales (contact@riseunion.io) for reference introductions.

Beijing RiseUnion Technology Co., Ltd. is a National High-tech Enterprise, Beijing Specialized SME, and chair of the AIIC Compute Pooling Working Group. We led the drafting of the Heterogeneous Compute Virtualization and Pooling System Requirements industry standard and co-authored the Diverse Computing Optimization Action Plan. Certifications include ISO 9001, ISO 27001, ISO 20000, and CMMI3, with 10+ software copyrights and patents. Products are deployed in MLPS Level 3 environments at multiple finance and government customers.
Learn more: About RiseUnion

RiseUnion is one of the core contributors to HAMi (Heterogeneous AI Computing Virtualization Middleware), a CNCF Sandbox project and the most active GPU virtualization open source project in China. We continuously contribute upstream and participate in project governance, while the enterprise-grade features (multi-GPU virtualization, domestic chip support, production-grade HA) are productized in Rise VAST Enterprise. Beyond HAMi, the team also contributes to Volcano, Kubernetes, and other upstream communities.
Further reading: HAMi v2.8.0 Release · HAMi Source Code Analysis

RiseUnion has raised tens of millions of RMB in funding (Plum Ventures sole investor) and received strategic investment from 4Paradigm. Rise VAST is co-developed with 4Paradigm as an enterprise AI compute management product. The company continues heavy R&D investment to maintain technical leadership in AI compute management and scheduling.