Choose the right deployment path for running a Livepeer orchestrator - solo, pool node, O-T split, siphon, and dual mode workload configuration.
Key Decisions for Orchestrator Setup
Setup Type - On-chain or off-chain
Setup Path - What software is installed
Operational Mode - Whether you control and handle all operational requirements or delegate them
Workload Mode - What compute job workloads the node processes
This page acts as a guide to finding the Orchestrator setup path that matches your operational aims.covers the deployment options available for Orchestrators by category and their focus:
A pool node is not a pool operator. A pool node joins someone else’s pool and contributes GPU compute. A pool operator runs the orchestrator that accepts external workers. These are different deployment types with different requirements.
The standard path. A single go-livepeer process on one machine handles protocol operations, job routing, and GPU work.The operator controls: pricing, stake, workloads, reward calling, uptime - everything.
A GPU-only process connecting to a pool operator’s orchestrator. No staking, no on-chain registration, no protocol management.The operator controls: which pool to join, GPU hardware, when to switch.The pool controls: registration, staking, pricing, reward calling, payout schedules.
Orchestrator and transcoder as separate processes on separate machines. The orchestrator handles protocol (no GPU). The transcoder handles GPU work. Connected by a shared secret.The operator controls: everything, but responsibilities are split across machines.
A secure machine runs OrchestratorSiphon (Python) for keystore, reward calling, and on-chain operations. A GPU machine runs go-livepeer in transcoder mode. Reward calling continues even when the GPU machine is down.The operator controls: everything, via OrchestratorSiphon on the secure machine.
Deployment type and workload mode are independent decisions. Any deployment type above can run any workload mode below.Dual mode is the most common production configuration. NVENC/NVDEC (video) use dedicated silicon that does not compete with CUDA cores (AI). Both workloads share VRAM. A 24 GB GPU supports video transcoding alongside one warm AI model.For full dual mode setup instructions, see .For a detailed breakdown of all AI pipeline types, VRAM requirements, and demand data, see .
