Blender cloud rendering on GPUs: setup, cost per frame, pitfalls

Why per-second billing fits rendering

Render load is the least steady workload in computing — nothing for days, then 400 GPU-hours before a deadline. Owning hardware for that peak means paying for idle; classic farms price the peak in. Per-second rental inverts it: 8 instances × 1 hour costs exactly 1 instance × 8 hours ($1.048 on interruptible 4090s), so parallelism is free and idle costs zero.

Setup: template + volumes

powergpu volume create --name scene --size 60 --region eu-west-1
powergpu volume create --name frames --size 100 --region eu-west-1

# stage the packed project once (any instance, or the console uploader)
powergpu launch --gpu rtx-4090 --template blender --volume scene:/scene
scp shot.blend root@i-…:/scene/

# test ONE frame before the farm — always
blender -b /scene/shot.blend -o /frames/f#### -F OPEN_EXR -f 40
# Fra:40 … Time: 03:41.20 (Saved: /frames/f0040.exr)

That single timed frame is your whole cost model: seconds-per-frame × frames × rate ÷ 3600. Never size a farm without it.

Turning instances into a farm

# 8 workers, frame ranges split by stride — no scheduler needed
for i in $(seq 0 7); do
  powergpu launch --gpu rtx-4090 --type interruptible \
    --template blender --volume scene:/scene:ro --volume frames:/frames \
    --env CMD="blender -b /scene/shot.blend -o /frames/f#### \
               -F OPEN_EXR -s $((1+i)) -e 250 -j 8 -a"
done
# each worker renders frames i, i+8, i+16, … — an interruption loses
# at most the frame in flight; -j strides make re-runs idempotent

Interruptible is the right mode because frames are idempotent queue items: a paused worker resumes its stride and re-renders one frame at half price. The Python SDK version of this loop (with auto-retry) is twelve lines.

Costing a shot honestly

Scenes over 24 GB (volumes, fur, 8K textures) move to the 48–96 GB workstation cards — the rendering playbook ranks them; the RTX PRO 6000 WS at $1.097/hr holds production scenes whole.

The pitfalls that triple render bills

1. Rendering PNG instead of EXR to "save space" — then re-rendering for the grade. Storage is $0.08/GB/mo; re-renders are GPU-hours. Always EXR.
2. Unpacked assets — workers render magenta placeholders for an hour before anyone looks. Pack resources, and eyeball frame 1 from every worker before walking away.
3. Per-frame instance churn — booting an instance per frame pays the 30-second boot 250 times. Workers take strides; boot once per worker.
4. Forgetting the farm — -a exits when the stride finishes, but the instance keeps billing its disk until destroyed. End every farm script with destroy (or use the SDK context manager that cannot forget).