A purpose-built GPU OS
for real-time inference
Not a general-purpose scheduler — an operating system for inference fleets
Models and latency budgets are the primitives. Compiled plowrt plans are placed across mixed NVIDIA and AMD fleets ahead of time, so at serve time the fleet has nothing left to decide.
General-purpose schedulers weren't built for token deadlines
Kubernetes was built to schedule containers; Slurm to schedule batch jobs. To both, a GPU is a countable device and a model is opaque cargo. A production inference fleet serves many models on many GPUs under tight latency budgets — and real-time workloads such as voice make the budgets explicit: a reply that streams late is a reply that failed. Running that through a general-purpose layer that assigns work as it arrives recovers utilization, but it reintroduces at fleet scale exactly what host orchestration causes inside a single GPU: decisions on the critical path, queueing variance, and tail latency that inherits every scheduling accident.
Static assignment avoids the jitter but wastes capacity; dynamic scheduling recovers capacity but sells the tail. Real-time serving needs a third option: an operating system for the fleet, built purely around inference.
A GPU OS: models as processes, budgets as the contract
The Orchestrator treats the fleet the way an operating system treats a machine — except its processes are models. Plow compiles a model into a fixed execution plan a GPU runs without consulting the host; the Orchestrator is being built to admit those plans against declared latency budgets, place them across GPUs, nodes, and vendors deliberately, ahead of time, and supervise them with measured telemetry — so that at serve time the fleet, like the GPU, has nothing left to decide.
A GPU OS where the model is the first-class unit
declare a model and its latency budget; the OS admits it, places plowrt plans onto GPU and CPU nodes ahead of time, and supervises them — where Kubernetes schedules containers and Slurm schedules batch jobs, this schedules models
It is designed around four commitments:
- SLO-aware placement. Models are packed onto GPUs against declared latency budgets, not best-effort queues — predictability is the contract, and admission control enforces it.
- Mixed fleets, one plane. NVIDIA and AMD nodes managed from a single control plane, because plow's plans already run natively on both.
- On-prem first. Bare metal, private cloud, sovereign and air-gapped deployments — the same Rust-first, no-Python-serving-stack posture as the engine.
- Measured feedback. Fleet telemetry flows through InferSight; placement decisions are calibrated by measurement, in the same spirit as the engine's auto-tuner.