Product
GPU/CPU Grid: distributed compute on idle machines
Plenty of CPUs and GPUs sit unused for most of the day. GPU/CPU Grid takes a compute workload, splits it into pieces and runs those pieces in parallel on available worker nodes, then releases the nodes when the work is done. It is built for workloads that can't wait in a queue.
The MVP already runs end to end, and it is the part of Symbola behind our 38x rendering result. Continuous real-time streaming is the next milestone.
How a job runs
- A workload comes in, for example a Blender frame range or a media stream.
- The control plane splits it into work ranges and assigns each one to an available worker node. Node selection is automatic.
- Nodes process their pieces in parallel. Data moves over end-to-end encrypted tunnels.
- Results are checked, reassembled and returned. Contributors are paid for the work their machines completed.
What fits
Anything that splits cleanly into independent parts. The more independent the parts, the closer the speed-up gets to the number of nodes.
| Workload | How it splits |
|---|---|
| Rendering and VFX | By frame range or tile |
| Simulations | By parameter set or region |
| Live video and media processing | By stream segment or frame |
| Streaming sensor and field data | By source or time window |
See how it applies to media production and simulations and edge workloads.
What's next
Next on the roadmap: job checkpoints and automatic reallocation when a node drops out, node health scoring, and support for real-time workloads. AI-assisted node selection and runtime prediction follow in 2028, trained on the operational data the grid collects.
See it run
The benchmark page has the full setup, the recorded run and the limits of the test.