Skip to main content

Architecture

How the pieces of a Lium provider setup fit together. Read this once to build a mental model — then proceed to Quickstart, Self-hosted provider setup, or Node Quickstart depending on the path you've chosen.

The two-component model​

A Lium provider setup has two distinct components:

  1. Provider — a lightweight CPU coordinator (no GPU). It signs with your registered subnet 51 hotkey, talks to validators, and pushes/pulls node configuration.
  2. Nodes — one or more GPU machines that perform actual rental workloads. Validators probe each node and renters connect to it.

A provider with no nodes earns nothing. Each node you bring online adds to your scoring surface and rental revenue.

Where the provider runs — your choice​

The CPU coordinator can run in two places:

OptionWho runs itSetup
Self-hosted providerYou, on a Linux box you provideSelf-hosted provider setup
Lium.io Central Provider ServerLium operates it for youToggle on in your Provider Portal profile

The choice is a per-account toggle in the Provider Portal profile page. You can flip back and forth — see Provider Configuration for the trade-offs and the toggle flow.

The nodes are always run by you, regardless of which path you pick for the coordinator.

How validators interact with your nodes​

Validators on subnet 51 do two things continuously:

  • Probe each node for the required sysbox-runc runtime, hardware specs, and synthetic workload performance. Nodes missing Sysbox are rejected — they earn no emission and cannot be rented. Setup is in Sysbox.
  • Score each provider based on those probes plus rental activity. Scores translate into TAO emission via Bittensor's standard subnet mechanism.

Rental income is separate from emission — renters pay you (in the platform's billing system) for actual GPU-hour usage. See Provider Portal → Payments for the payout view.

What goes on each component (component breakdown)

On the provider host (CPU coordinator)

  • Bittensor btcli for hotkey operations
  • The provider Docker container (from lium-io/neurons/miners/)
  • The registered subnet 51 hotkey (only the hotkey — never the coldkey)

On each node host (GPU)

Off-host (managed for you)

  • The Provider Portal — node registration, pricing, monitoring, payouts
  • The Bittensor network — subnet 51 emission, hotkey state
  • Optionally, the Lium.io Central Provider Server if you opted in

Supported GPUs​

What GPU models can I bring to this subnet?​

The validator scores a fixed allow-list of NVIDIA models — anything outside this list is rejected at machine-spec time. Within the list, only some models are admitted to the unrented pool (so they earn the unrented-emission share); the rest are operationally supported (you can register and rent them out) but receive 0 from the unrented-pool reward.

The live list — with each model's reference rental price — is available from the API:

curl https://lium.io/api/machines

Price limits apply to the reference price returned by this endpoint: the portal enforces a 0.5× floor and a 3× ceiling (HTTP 400 outside that range). The static list below is a snapshot; the API is authoritative. The validator source also documents each model's emission weight — see neurons/validators/src/services/const.py → GPU_MODEL_RATES.

Datacenter — Hopper / Blackwell​

B300 SXM6 AC, B200, H200, H200 NVL, H100 80GB HBM3, H100 NVL, H100 PCIe, H800 80GB HBM3, H800 NVL, H800 PCIe.

Datacenter — Ampere & inference​

A100 80GB PCIe, A100-SXM4-80GB, A10 Tensor Core, T4 Tensor Core, Tesla V100 Tensor Core, Tesla P100, Tesla P40, Tesla M40.

Workstation — Ada, Ampere, Blackwell​

RTX 6000 Ada, RTX 5880 Ada, RTX 5000 Ada, RTX PRO 6000 Blackwell (Server / Workstation Editions), RTX PRO 5000 Blackwell, RTX PRO 4000 Blackwell, RTX A6000, RTX A5000, RTX A4500, RTX A4000, RTX A2000, L40S, L40, L4.

Workstation — older Quadro / TITAN​

Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 5000, Quadro P4000, TITAN RTX, TITAN V, TITAN Xp.

Consumer — RTX 50-series​

RTX 5090, RTX 5080, RTX 5070 Ti, RTX 5070, RTX 5060 Ti, RTX 5060.

Consumer — RTX 40-series​

RTX 4090, RTX 4090 D, RTX 4080 SUPER, RTX 4080, RTX 4070 Ti SUPER, RTX 4070 Ti, RTX 4070 SUPER, RTX 4070, RTX 4060 Ti, RTX 4060.

Consumer — RTX 30-series​

RTX 3090 Ti, RTX 3090, RTX 3080 Ti, RTX 3080, RTX 3070 Ti, RTX 3070, RTX 3060 Ti, RTX 3060, RTX 3060 Laptop, RTX 3050.

Consumer — RTX 20-series & GTX 10/16-series​

RTX 2080 Ti, RTX 2080 SUPER, RTX 2070 SUPER, RTX 2060 SUPER, RTX 2060, GTX 1660 Ti, GTX 1660 SUPER, GTX 1660, GTX 1080 Ti, GTX 1080, GTX 1070 Ti, GTX 1070, GTX 1060.

Rented vs unrented pool eligibility

Every model on the list is eligible for rental revenue (the customer-paid stream). Only a subset earns the unrented-emission share — concretely, the models with a non-zero entry in GPU_MODEL_RATES. Models marked 0.0 in that dict (e.g. RTX 5080, RTX 4080, A10, T4, RTX A2000, every model older than the RTX 30-series) still mine and rent normally; they just don't accrue unrented-pool weight when idle. For live per-model reference prices, query https://lium.io/api/machines.

CVM-eligible GPUs

For Confidential VM (CVM) nodes, only Hopper (H100, H200) and Blackwell (B200, GB200) are supported — consumer and workstation cards do not implement NVIDIA Confidential Computing.

Where to go next​