Skip to content
Cindergrid

Node classes

Updated 2 November 2026. Applies to scheduler 2026.11 and sync client 2.8.

Every job runs on one node class. Pick by memory first, speed second: a frame that runs out of memory is slower than one that runs on fewer, larger nodes.

ClassCoresRAMGPUScratchRateRegions
Kiln C6464256 GB-2 TB NVMe$0.0055 / core-hrall 7
Kiln C128128512 GB-4 TB NVMe$0.0049 / core-hrall 7
Kiln M96961.5 TB-4 TB NVMe$0.0068 / core-hreu-north, eu-central, us-east, us-west
Arc G11664 GB1 x 24 GB1 TB NVMe$0.62 / GPU-hrall 7
Arc G224128 GB1 x 48 GB2 TB NVMe$1.10 / GPU-hrall but ap-southeast
Arc G496768 GB4 x 80 GB4 TB NVMe$2.35 / GPU-hreu-north, us-east, us-west

Choosing

  • Arnold, V-Ray CPU, Corona, Karma CPU: C64 for most shots; C128 when a frame uses more than 180 GB or when you want fewer nodes to babysit.
  • Out-of-core scenes above 400 GB: M96.
  • Cycles GPU, Redshift, Eevee Next: G1 when geometry and textures fit in 20 GB, G2 up to 44 GB.
  • Karma XPU, Unreal Path Tracer: G4, which keeps all four GPUs in one memory pool.

Benchmark scenes

We keep four public scenes to compare classes. Figures are minutes per frame at 1080p, 512 samples, median of 12 runs on scheduler 2026.10.

SceneC64C128G1G2G4
Kitchen interior (Cycles)11.26.13.42.50.8
Hair and cloth (Arnold)19.810.6---
Pyro plume (Karma)34.018.2-9.82.9
City flythrough (Unreal PT)--7.34.91.6
GPU node illustration

Isolation

A node serves one project at a time. Disks are cryptographically erased between jobs and images are rebuilt on the first Tuesday of every month. Nodes in a job cannot reach each other except through the scheduler, and have no route to the public internet unless you enable it for asset fetching.