Rendering network using crowdsourced GPU

The graphics rendering service consists in a network of decentralized machines called nodes which are of 3 kinds: manager, worker and verifier. The managers are dedicated machines of Inferix while verifiers and workers are machines joined by GPU owners. The number of workers is normally much larger than the number of managers and verifiers.

Figure 1

Rendering flow

A typical rendering session contains several steps as shown in Figure 1 and explained below:

1. A user creates a rendering job request using the Inferix's plugin for client, this job uploads user's scene data to some manager.

2. The rendering task controller of the manager receives the rendering job request, then

   1. splits it into multiple rendering tasks, each task consists of the scene data and several parameters: range of frames to be rendered, output format, etc.

   2. generates corresponding verification keys and sends these keys to the verifying task controller.

   The rendering tasks will be assigned to workers and the verification keys will be sent to the verifying task controller.

3. Receiving a rendering task, a worker renders the included scene using the parameters given by the task. When it finishes, it saves the rendered frames to a decentralized storage, then notifies the manager by a message containing a unique URL to the result.

4. The verifying task controller of the notified manager receives the notification then creates a verification task; this task will be assigned to a verifier.

5. Receiving a verification task, a verifier

   1. checks the authenticity of the corresponding rendered frames, then

   2. notifies the manager about the verification result.

6. If the rendered frames pass the verification, then the manager notifies the user by a message containing the URL to the rendered frames. Otherwise, these frames are rejected.

7. The user downloads the frames from the storage and manually confirms whether they meet the expectation, if they do not then the user sends a bad result claim to the manager.

The managers synchronize a database of rendering and verification tasks. That makes the rendering service being both logically and physically decentralized: a graphics scene can be simultaneously rendered by different workers and later checked by different verifiers, the machines of workers and verifiers can be also located at different geographical locations.