Documentation

How GPUFarm works

GPUFarm is a marketplace for GPU compute. Hosts connect machines and group them into farms; customers submit containerized jobs; a scheduler matches each job to compatible hardware; the workload runs offchain in a sandbox; the result is verified and payment settles from escrow on Robinhood Chain.

Run a hostInstall the daemon, pair it to your wallet, pass the benchmark and go online.REST APICreate jobs, fund escrow, stream logs and download results with scoped API keys.TypeScript SDKgpufarm.jobs.create(…), poll, download — and fund the escrow with a viem wallet.Workload securitySandbox flags, network policy, signed artifact URLs and host-key isolation.Trust modelWhat is onchain, who verifies compute, disputes, and what each key can and can't do.MarketplaceLive, benchmark-verified capacity with per-GPU-hour prices.

Architecture#

Compute never runs onchain. The chain holds escrow, receipts and payouts; everything else is offchain and labeled honestly about what it proves.

1Customer
Configures a job and sees a live price estimate with its assumptions.
2Escrow
Funds the job's maximum cost into ComputeEscrow on Robinhood Chain.
3Scheduler
Matches GPU class, VRAM, count, region, load, price and reputation — never cheapest-only.
4Host
Accepts the 10-second offer, pulls the image and inputs, runs the sandboxed container.
5Verification
Server re-hashes the output, checks GPU telemetry, re-computes deterministic workloads.
6Settlement
Verifier signs the completion; after the challenge window (or your approval) payment releases.

Roles#

Role	Does	Starts at
Customer	Rents GPUs: configures jobs, funds escrow, downloads results, approves or disputes.	/marketplace
Host	Connects one or more GPU machines with the daemon, sets prices and policy, earns per verified GPU-second.	/host
Farm owner	Groups machines into a public compute pool with a name, region, prices and an optional commission.	/farms

Jobs#

A job is one container: an image, a command, GPU requirements (count, minimum VRAM, optional model allowlist), a max runtime, output rules and a network policy (none by default). V1 runs every GPU of a job on one physical machine — there is no cross-host distributed training.

Workload types hosts can accept or decline:

AI Inference inferenceImage Generation image_generationVideo videoRendering renderingBatch Compute batchFine-Tuning fine_tuningDevelopment developmentCustom Container custom

Every job moves through these states. Failed compute is never charged.

State	Meaning
CREATED	Configured and priced. Fund the escrow to queue it.
FUNDED	Payment is locked in the escrow contract on Robinhood Chain.
QUEUED	Waiting for compatible capacity.
MATCHING	The scheduler is offering this job to a compatible host.
ASSIGNED	A host accepted the job and reserved the GPUs.
DOWNLOADING	The host is pulling the container image and inputs.
RUNNING	The workload is executing on the assigned GPUs.
UPLOADING	The host is uploading outputs, logs and metrics.
VERIFYING	GPUFarm is checking the result hash and execution telemetry.
COMPLETED	Result delivered. Payment settles after the challenge window or your approval.
FAILED	The job could not complete. You are not charged for failed compute.
DISPUTED	Escrow is frozen while an arbitrator reviews logs, telemetry and hashes.
CANCELED	Canceled before completion.
REFUNDED	The escrowed payment was returned to the customer.

Verification, briefly#

Result verified — a deterministic verifier re-computed the output (V1: the gpufarm/selftest:1 workload).
Execution verified — the output hash matches the host's signed receipt and GPU telemetry confirms the assigned GPUs were working.
Host reported — integrity holds, but execution could not be independently confirmed. Use the challenge window.

The full picture — who signs what, how disputes work and what a compromised key could do — is on the trust model page.