Rendering on GPU Rental Platforms

Why Rent GPUs on Demand

For individual creators and independent studios, local workstations have limited compute power, and purchasing multiple high-end rigs is capital-intensive.

Traditional commercial render farms charge per-frame or credit markups, making costs unpredictable while offering limited live visibility into progress. By renting modern cloud GPU instances (such as RunPod, Vast.ai, or AutoDL), you can spin up single or multi-card RTX 4090 / 3090 instances, pay by the second or hour, and destroy them immediately when rendering finishes—often cutting total compute expenditure by 70% to 80% compared to legacy farms.

Combined with danio's single-command joining and resilient lease scheduling, cloud nodes automatically take on workloads just like local workstations in the next room.

Hardware & GPU Selection Guide

Blender Cycles heavily relies on CUDA / OptiX ray-tracing cores and high core frequencies. When picking hardware, we recommend consulting the official Blender Open Data benchmark leaderboard:

• Speed flagship recommendation (RTX 5090 / RTX 4090): Dominating the top tier of Blender Open Data. The RTX 5090 (32 GB VRAM) and RTX 4090 (24 GB VRAM) deliver unmatched OptiX ray-tracing throughput at a fraction of the cost of enterprise compute accelerators (typically $0.35–$0.89/hr), offering unbeatable price-to-performance.
• Massive scenes & heavy VRAM recommendation (RTX 6000 Ada / RTX Pro 6000): For large-scale architectural projects, heavy 8K UDIM texture sets, or dense volumetric simulation caches, choose professional 48 GB workstation cards (such as the RTX 6000 Ada / Pro 6000 series) to completely avoid CUDA Out of Memory crashes.
• High-value alternative (RTX 3090): Features a generous 24 GB VRAM with ultra-low rental rates (~$0.20–$0.30/hr), ideal for overnight batch rendering and long shot sequences.
• Avoid legacy compute cards: Older cards (such as P40 or T4) and pure CPU instances lack modern RT cores and render many times slower.

Recommended Platforms & Comparison

Prerequisites

1. S3 Object Storage Workspace: Cloud nodes cannot reach local SMB network shares. The workspace must use S3 storage. We strongly recommend Cloudflare R2 for its $0 egress fees, avoiding unexpected transfer bills when workers pull assets concurrently.
2. Public HTTPS URL for Host: Cloud nodes must reach your host via a public URL (such as https://danio.yourdomain.com), see Remote Devices.
3. Linux Blender Resource Pack: Most rental platforms run Linux (Ubuntu). In the host under Workspace details → Resource packs, import a Linux version of Blender (see Resource packs); joined nodes will download it automatically.

Step 1: Generate Joining Command on Host

On the danio host, go to Devices → Add device:

1. Enter a unified Name prefix, e.g., cloud-gpu.
2. Set Permission to None: Cloud render nodes only need to pull chunks and upload rendered fragments. They do not need access to view other projects or configure the workspace.
3. Turn on Multiple devices: A single token can be shared across multiple spun-up nodes, automatically numbering them cloud-gpu-01, cloud-gpu-02...
4. Keep Render checked.
5. Set Download danio to From official source (nodes download via GitHub CDN, which is fast and spares your host's upstream bandwidth).
6. Configure Local Scratch Cache (Critical): Open Extra settings and specify the node's local NVMe scratch path:
   • RunPod: /workspace/cache
   • Vast.ai: /root/cache
   • AutoDL: /root/autodl-tmp/cache (do not place on /root/autodl-fs network drives)
7. Click Generate token, select Linux as the system, and copy the full command.

Step 2: Launch and Join on Rental Platforms

Deploying on RunPod

1. In the RunPod dashboard (referral link), navigate to Pods and click Deploy.
2. Select your GPU (e.g., 1x RTX 4090), and pick Community Cloud (budget) or Secure Cloud.
3. Choose a template image (e.g. runpod/pytorch:2.4.0-py3.11-cuda12.4.1-devel-ubuntu22.04 or standard Ubuntu).
4. Expand Container Disk (recommend at least 50 GB to store unpacked textures and job directories).
5. Launch the Pod. Once Running, click Connect → Start Web Terminal.
6. Paste the danio join command copied from your host and press Enter.
7. The installer automatically installs dependencies, registers the node, and starts danio serve in the background.

Deploying on Vast.ai

1. Log into the Vast.ai dashboard (referral link) and open the Search filter.
2. Filter for RTX 4090, and optionally check Datacenter.
3. Set Disk Space to at least 50 GB.
4. Select the official template nvidia/cuda:12.4.1-runtime-ubuntu22.04.
5. Click Rent. Wait for the instance to become active under Instances.
6. Click the SSH button on the instance, copy the SSH command, and connect from your local terminal.
7. Run the danio join command.

Deploying on AutoDL

1. Log into the AutoDL console and rent an RTX 4090 instance.
2. Select Ubuntu 22.04 with CUDA 12.x.
3. Open JupyterLab terminal or connect via SSH.
4. Run the join command. Danio will automatically utilize the /root/autodl-tmp local data volume.

Step 3: Verification & Rendering

Once connected, return to the host's Devices page:

• The new machine will appear as cloud-gpu-01 with status Online.
• Open device details: Capabilities will display Cycles configured as OptiX with your GPU model detected.
• If tasks are waiting in the queue, the node will immediately switch to Preparing (downloading Blender and missing assets) and then begin Rendering.

Fault Tolerance & Preemption Protection

When using on-demand or spot/interruptible instances, nodes can occasionally disconnect or get reclaimed by providers:

• 2-Minute Lease Protection: danio grants a 2-minute dynamic lease for every chunk. If a server abruptly drops offline, the lease expires and the host reassigns uncompleted frames to other active workers.
• Completed Frame Protection: In the task details, click any active chunk and select Recall this chunk…. Checking "Keep rendered frames" encodes all completed frames up to that moment into a fragment immediately, handing only remaining frames to another node.

Teardown & Clean Up

When all shots finish rendering and the film is joined:

1. Terminate/Destroy the instance in your provider's dashboard (RunPod / Vast.ai) to halt billing.
2. In the danio host Devices page, open the offline device's details card.
3. Click Block at the bottom, then hold Remove device for 2 seconds to purge it from the roster.