A farm that behaves like part of your pipeline
Cindergrid is made of four things: a transfer layer for very large scenes, a scheduler, the nodes, and a place to watch it all happen. Each can be used on its own through the API or CLI.
Transfer
Scene data is the slow part of cloud rendering. We built the client around it.
The sync client, cgsync, splits files into 8 MiB chunks, hashes each with xxh3 and asks the farm which chunks it already holds. Only missing chunks travel. A second version of a scene that changed 3% uploads 3%.
Uploads use up to 32 parallel streams (8 on Spark, 16 on Studio, 32 on Pipeline), pause and resume across sleep and network changes, and keep working through a flaky hotel connection. On a clean 10 Gbit/s uplink we routinely see 1.1 GB/s sustained to eu-north.
- Alembic, VDB, bgeo.sc and USD caches are recognised and uploaded in dependency order.
- Texture and asset paths are rewritten on the farm; your scene file is never modified locally.
- Scene storage is encrypted at rest with per-project keys. Frames are encrypted in transit and at rest.
Scheduler
Deadline-aware rather than first-come, first-served.
Tell Cindergrid when you need a shot and it sizes the job: how many nodes, in which region, at what cost. Priority within a plan is set per job, and Studio and Pipeline jobs jump ahead of Spark jobs only when capacity is short, which in 2026 has been under 1% of the time.
- Frame chunking. Chunk size is chosen from a three-frame probe so short frames are not drowned in startup time.
- Tile splitting. A single 16K frame can split across up to 64 nodes and is stitched on the last one.
- Dependencies. Simulate, then render, then comp. A job waits for its parents and inherits their caches.
- Cost caps and tags. Cap a job at a dollar figure; tag by show or department and export usage per tag as CSV.
- Automatic retries. Failed frames retry twice on different nodes, then pause the job and tell you why.
Nodes
Same images in every region, rebuilt monthly.
Every node boots a clean image with the DCC, renderer and plugin versions pinned by your job. Nothing is shared between customers: a node serves one project at a time and is wiped before the next. Local NVMe scratch holds working data and is destroyed with the node.
See node classes for cores, memory and per-class benchmark scenes, and renderer licences for how Arnold, V-Ray and Redshift are handled.
Monitoring and review
See what is running, what it costs and what it looks like.
The job page shows frames as they finish with a per-pass viewer, a cost meter that updates every few seconds and a log tail per frame. Shot supervisors can approve a look or stop a job from a phone.
Example: studio queueupdates live
Webhooks fire on job start, frame failure, cost-cap reached and completion. The same events are available from cgsync watch for studios that prefer a terminal. A read-only status board for the whole studio can be shown on a wall display.
Compare plans or read the changelog to see how fast this is moving.