Heterogeneous Foundation
Heterogeneous
Foundation
A software platform for hyperscale AI training and inference
with a built-in Agentic operations system
Across architectures, regions, and ownership models
Optimizing Heterogeneous Chips to Maximize Resource Utilization
Native pooled management and intelligent scheduling for 16 mainstream AI accelerators
Thriving hardware ecosystem
Intelligent pooling and scheduling
End-to-end real-time observability
Automated operations fault handling
16
Mainstream Chips
85%+
Improved Resource Utilization
50+
Core Metrics Monitoring
85%
Faults auto-self-healing
Capabilities
Comprehensive chip ecosystem: Full support for major chip families, including NVIDIA, AMD, Huawei Ascend, Iluvatar, MetaX, Biren, and more.
Hardware abstraction: A unified resource abstracts away hardware differences, allowing upper-layer applications to remain agnostic to underlying accelerator types.
Unified interface compatibility: Standardized interfaces for resource requests, monitoring, and management, with compute capacity, memory, and interconnect bandwidth quantified in a consistent manner.
Automatic resource pooling: Supports automatic pooling across multiple dimensions, including accelerator type, geographic region, and workload scenarios.
Dynamic resource allocation: Dynamically reallocates resources across pools to adapt to workload changes, with automatic consolidation of fragmented resources to reduce inefficiency.
Multi-strategy scheduling algorithms: Supports a wide range of intelligent scheduling strategies, including topology-aware scheduling, bin packing, gang scheduling, priority adjustment, and preemptive scheduling.
Multi-dimensional data collection: Comprehensive data collection across hardware-level, task-level, and cluster-level metrics, unified within a single monitoring system.
Multi-level intelligent alerting: Alert prioritization, aggregation, and suppression based on contextual scenarios, reducing invalid alerts by more than 80% and supporting real-time, multi-channel notifications with closed-loop handling.
Cross-platform collaboration: Alerts and signals are integrated across cluster management, operations platforms, and training/inference services, breaking down information silos from infrastructure layers to application layers.
AI-driven anomaly detection: Predicts and identifies a wide range of anomalies, including performance degradation, thermal issues, and slow nodes, with automatic root cause analysis.
Intelligent health checks: Monitors the health of hardware, networks, and storage systems, with periodic automated probing and validation tasks.
Automatic fault handling: Automatically isolates faulty nodes and migrates affected workloads, with support for automatic restarts and configuration rollbacks.
Why Us
Use Case
A Research Institute
To build an autonomous and controllable heterogeneous compute infrastructure, the institute needed to introduce multiple domestic AI accelerators in phases while continuing to operate existing imported hardware, ensuring a smooth transition for ongoing research workloads.
Fragmented chip management: prevented unified management of imported and domestic chips.
High migration costs: Workload migration required repeated adaptation for each chip type, resulting in high engineering overhead.
Complex troubleshooting: Troubleshooting required coordination across multiple teams and vendors, leading to long resolution chains and low efficiency.
Inefficient in-house integration: Self-integration involved long development cycles, slowing research progress and forcing compromises on functionality.
Slow deployment progress: Inefficient scheduling and resource allocation delayed adoption, increasing friction with internal research teams.
Unified management: A single software suite manages all chips, providing unified monitoring and pooled resource management.
Multi-tenant isolation: Multiple departments are isolated by budget-based quotas, ensuring independent and non-interfering operations.
Reduced migration costs: Pre-integrated platform support significantly reduces migration effort, requiring only minimal application-level changes.
Higher operational efficiency: End-to-end observability enables rapid fault localization, substantially improving operations efficiency.
Intelligent scheduling: Flexible, policy-driven scheduling enables resource sharing and borrowing, increasing utilization and reducing wait times.
A Provincial Computing Center
The center aggregates compute resources from multiple regions and vendors, delivering unified computing services to a wide range of provincial business units.
Challenging budget allocation: Budgets from different departments could not be proportionally allocated across heterogeneous accelerator types, leading to frequent disputes over resource selection.
High migration dependency: Workloads were tightly bound to specific chip models, requiring bilateral coordination with vendors and incurring high adaptation costs.
Lengthy onboarding cycle: Integrating new compute suppliers was complex, with onboarding cycles lasting several months and resulting in persistent resource shortages.
Flexible quotas: Budget-based multi-tenant isolation lets departments allocate quotas flexibly and track usage in real time.
On-demand scheduling: Departments can specify required accelerator types, while the pooled scheduling system automatically provisions workloads without additional adaptation.
Rapid onboarding: New hardware is automatically standardized and onboarded by the platform, dramatically improving supply efficiency and overall user experience.
Infinite Computing, Accelerating the Future of AGI
Contact us for customized AI infrastructure solutions
