One-line positioning
For distributed AI inference tasks in satellite constellations, build a ground-based simulation and validation environment combining “multi-satellite computing + dynamic networks + real model execution” to quantitatively validate distributed inference strategies before algorithms go onboard and systems are integrated.
Solution overview
As satellite constellations expand and remote sensing, communication and in-orbit computing tasks grow, AI inference is moving from independent execution on single satellites toward multi-satellite and space–ground collaboration. These tasks are constrained by CPU, GPU, memory and storage resources and tightly coupled with inter-satellite links, satellite-to-ground links, visibility windows, link switching and task arrivals.
Existing validation methods separate computing from networking: general cloud or edge computing test platforms usually assume relatively stable networks, making it difficult to represent satellite topology, visibility windows and continuously changing inter-satellite / satellite-to-ground links. Traditional satellite network simulation tools calculate topology, routing and link performance but usually do not run real AI models, so they cannot accurately reflect GPU, memory and cross-node execution overhead. Algorithm experiments on real satellites or complete flight systems are costly, slow and difficult to control and repeat.
The solution provides integrated ground-based hardware and software validation infrastructure, defined as: multi-satellite logical environment construction → execution on domestic heterogeneous computing resources → time-varying space–ground network simulation → closed-loop algorithm integration → full-process quantitative evaluation.
Core capabilities
① Joint multi-satellite computing and network simulation
- Time-sequenced simulation of satellite positions, ground station positions and all network link connections based on constellation orbit parameters
- Automatically update ISL and satellite-to-ground access link connectivity with simulation time, reproducing dynamic topology switching, periodic link connections/disconnections and changing access relationships caused by high-speed LEO satellite motion
- Fully support inter-satellite and satellite-to-ground access link simulation, with optional extension to terrestrial transport network segment modeling to build an integrated satellite–ground–ground end-to-end simulated network
- Synchronize network state and AI task execution state on a unified simulation timeline
② Inline impairment simulation
- Accurately simulate link degradation caused by rain attenuation, ionospheric disturbance, atmospheric obstruction, satellite beam handover and satellite-to-ground access switching
- Support fine-grained customized network impairments: bandwidth limits, fixed / dynamic delay, random jitter and on-demand packet loss on any network link
- Support lightweight multi-instance deployment using Docker containers for distributed impairment simulation of large numbers of parallel inter-satellite / satellite-to-ground links
- Visually edit and orchestrate multi-stage, multi-type link impairment actions on a unified simulation timeline
③ Heterogeneous computing resource simulation and real execution
- Rather than allocating a separate physical computer to each satellite computing node, use virtual machines and containers on a domestic physical computing infrastructure to create multiple logical satellite computing nodes
- Each node maintains separate resource views for CPU quota, available GPU capability, memory, storage and I/O, running an independent task agent and real AI workloads
- Performance profiling: run operator, model, memory / GPU memory bandwidth, I/O and concurrency benchmarks on physical platforms for specified computing payload chips or boards to create versionable performance profiles
- Profile injection: parameterize measured operator latency, model execution latency, resource usage and concurrency characteristics and inject them into each logical satellite node’s resource model; different nodes may load different chip profiles or use different resource quotas
- Real execution: mainstream example models, task agents, scheduling interfaces and metric collection components run in virtual machines or containers, generating real computation, queueing, synchronization and data transfer behavior
- Closed-loop calibration: feed real task execution results back into the metric repository to calibrate resource models, supporting reproducible comparisons of the same task across payload profiles, node configurations and scheduling strategies
④ Distributed AI inference task execution
- Describe AI inference tasks as task graphs with computation and data dependencies, supporting single-satellite baselines and model partitioning across satellites
- For Transformer inference, partition consecutive layer groups across satellite computing nodes: Prefill processes all input tokens in parallel and builds the KV Cache; Decode reuses the KV Cache for autoregressive token-by-token execution
- Run each model segment in reality while simulating inter-satellite link transmission, recording stage execution, queueing, synchronization waits and communication overhead
- Configure model partition points, node mappings, input / output lengths, batch size, task priority and latency targets
- Start, pause, migrate, terminate and query tasks, with consistency checks between single-node and distributed execution results
⑤ Scheduling algorithm integration
- Read system state through gRPC, REST or SDK; external algorithms output decisions on task partitioning, placement, resource allocation, execution order, migration or communication paths
- Run the scheduling loop as “state collection → algorithm invocation → decision validation → command execution → metric feedback”, with a configurable default control period of 10 seconds
- Validation objective: compare how strategies affect end-to-end latency, timeout rate, resource utilization, communication overhead and execution stability under the same constellation scenario, resource state and AI workload
Typical application scenarios
| Scenario | Description |
|---|---|
| Cross-satellite model partitioning and pipeline inference | Divide an AI model into consecutive or parallel stages deployed on different satellite nodes, transferring intermediate features or partial results between nodes; record end-to-end latency, stage waiting time, throughput and resource utilization |
| Multi-satellite task placement and dynamic migration | Select execution nodes based on satellite visibility, node resources and task service levels; trigger migration or redeployment when node load, links or target ground stations change; compare task completion rate and tail latency across strategies |
| Space–ground collaborative inference | Place inference stages on satellite nodes, ground stations or terrestrial computing nodes according to data location, model stage, onboard resources and downlink windows; measure the combined end-to-end impact of intermediate data downlink, space–ground communication waits and terrestrial execution |
| Intelligent remote sensing image processing | Benchmark task type |
Four operating modes
| Mode | Main use | Characteristics |
|---|---|---|
| Logical simulation mode | Large-scale constellation and network scenario validation | Focus on scaling logical nodes and dynamic topology; AI tasks may use performance models |
| Real-time simulation mode | Scheduling control loop and online decision validation | Advance network and task state by wall-clock time, with algorithm and system running concurrently |
| Real task execution mode | Model execution and performance measurement on domestic GPUs | Execute AI tasks on actual hardware and collect inference latency, resource and migration data |
| Hybrid validation mode | Large background constellation plus a few real execution nodes | Balance scale, realism and hardware cost |
Key technical specifications
| ID | Metric | Specification | Validation method |
|---|---|---|---|
| M-01 | Multi-satellite computing-network simulation scale | At least 10 independently configurable satellite computing nodes, scalable to at least 100 | Build scenarios and inspect node topology and CPU / GPU / memory / storage configurations |
| M-02 | Topology update period | No more than 15 seconds, configurable | Record time from scenario input changes to system topology state updates |
| M-03 | Real task concurrency | At least 10 concurrent inference task instances under agreed benchmark model and hardware configurations | Use agreed models and input data to inspect running task states and result records |
| M-04 | Network impairment parameter simulation | Configurable bandwidth limits, delay, jitter and packet loss; measured deviation of key parameters no more than 10% under agreed conditions | Compare configured values with end-to-end measurements |
| M-05 | Scheduling control period | Supports a configurable 10-second control period | Run multiple consecutive cycles and inspect state collection, algorithm invocation, decision validation and command execution |
| M-06 | Metric recording granularity | Periodic resource and network metric sampling at intervals of no more than 1 second, configurable | Inspect timestamps, event completeness and exported data |
| M-07 | Experiment reproducibility | Repeat the same fixed scenario at least 3 times, with key metric coefficients of variation no more than 10% | Repeat experiments with the same configurations and inputs and compare results |
| M-08 | System stability | At least 24 hours of continuous stable operation | No abnormal system exit or core function failure during long-duration stability testing |
| M-09 | Domestic GPU adaptation | Complete adaptation for 1 domestic GPU product | Correct real-model execution, with deployment, performance and compatibility reports |
Distributed inference evaluation metrics: end-to-end performance (task completion time, P95/P99 latency, timeout rate, time to first token and per-token latency), stage performance (execution time, queueing time and synchronization waiting time for each inference stage), communication overhead (intermediate data volume, inter-satellite / satellite-to-ground transfer time and retransmission overhead), computing efficiency (CPU / GPU utilization, GPU memory / memory usage and resource consumption per task), scheduling overhead (algorithm decision time, control period and policy update count), and result consistency (distributed versus single-node inference result consistency and accuracy changes).
Domestic platform adaptation
- Prefer domestic GPU products for the execution infrastructure, adapting drivers, compilation tools, runtime, performance monitoring tools and AI software stacks
- Determine domestic sourcing scope for CPUs, operating systems, container runtimes and storage/network components according to procurement lists and deployment environments
- Adaptation includes deployment, conversion, execution and correctness validation of typical AI models on domestic GPUs; benchmark establishment for model loading, inference latency, GPU memory usage, throughput and concurrent execution; resource isolation and performance interference validation under multi-process or multi-task concurrency
Deliverables
Integrated appliance (domestic GPU, CPU, memory, storage and network components, with preinstalled system software); multi-satellite computing-network simulation software; distributed inference runtime; algorithm integration interfaces (documentation, examples and baseline strategies); monitoring and visualization tools; domestic platform adaptation package; test scenarios and benchmark task sets; technical documentation; test reports.
Target users
- Distributed inference algorithm research teams
- Onboard task orchestration and computing-network scheduling research teams
- Domestic computing platform research and adaptation teams
- Satellite constellation system design and in-orbit computing planning organizations
Verifiable facts
- Complete system capabilities are available for multi-satellite computing-network simulation, heterogeneous computing resource virtualization and profiling, real AI task execution and distributed scheduling algorithm integration
- Complete interface designs are available for satellite orbit, ground station and application task data inputs
- Four operating modes are available (logical simulation / real-time simulation / real task execution / hybrid validation)
