Edge Microvisor Toolkit is available in several pre-configured versions that serve different purposes. Some are published as binaries, others are available from a custom build. This document will help you select the version that best suits your needs. To do so, check out:
| Version | Real Time | Stable Kernel | Next Kernel |
|---|---|---|---|
| Standalone (Immutable) | Available for opt-in | ✓ | ✓ |
| Developer Node (Mutable) | Optional | ✓ | ✓ |
| EMT for EMF | Available for opt-in | ✓ | ✓ |
| Bootkit | - | ✓ | – |
Run latency-sensitive workloads with guaranteed bounded jitter and repeatable
execution timelines across one or more hosts, maintainable under steady-state
and failure-recovery conditions.
Primary outcomes:
- Bounded end-to-end latency & jitter
- Repeatable scheduling windows under load
- Cross-host timing consistency for distributed stages
- Fast, predictable recovery without violating SLOs
Technology areas:
- PREEMPT_RT kernel
- Resource Director Technologies
- Intel GPU RT
- CPU & Scheduler Isolation
- Memory Determinism
- Network Determinism (TSN)
- Time & Clocks
Kernel patchsets (quilts):
Virtual Machines (Kubernetes, shared GPUs)
Run multiple virtual machines on Kubernetes that concurrently share one or more
physical GPUs, with predictable fairness, isolation, and policy-driven placement—using
a KubeVirt stack extended for GPU sharing.
Primary outcomes:
- Stable, repeatable GPU performance per VM under contention
- Hard/soft sharing policies (fair-share, priority tiers, or quotas)
- Safe isolation between tenants/VMs (memory, contexts, resets)
- Schedulable resources with clear admission signals (no surprise fails)
- Operational guardrails: health checks, graceful drain/eviction, rollback
Technology areas:
Kernel patchsets (quilts):
AI & Vision Systems
Enable AI inference and computer-vision workloads on edge nodes using Intel GPU
and NPU acceleration, exposing unified hardware-assisted pipelines through
standard APIs and user-space libraries.
Primary outcomes:
- Efficient execution of deep-learning and vision inference on-device without cloud dependency
- Unified GPU/NPU compute abstraction for developers (OpenVINO backend, media pipelines)
- Deterministic frame-rate and latency for multi-stream analytics workloads (e.g., camera ingest)
- Seamless integration with containers or pods, including dynamic device discovery and sharing
- Stable ABI/API interface across OS updates and driver versions
Technology areas:
You can create your own custom version of Edge Microvisor Toolkit by following the guide. You can also try and learn how to build your own solution and deploy it on edge.