JasmyLab compared which GPUs Blender render farms use and how much GPU memory a Cycles job gets, using each company's own pages, alongside the GPU that JANCTION Render, its Blender cloud GPU rendering service for AI agents, runs on. Most farms use cards with 24 to 48 GB. Every JANCTION Render job runs on a 96 GB NVIDIA RTX PRO 6000 Blackwell.
- On each company's own pages (checked November 16, 2026), Ranch Computing, iRender and Renderjuice list 24 to 32 GB cards such as the RTX 4090 and 5090; RenderStreet and BlendSwap list the 48 GB L40S.
- 96 GB cards are offered on GarageFarm's Medium and High tiers and by JANCTION Render, which has no tier on a smaller card.
- A Cycles scene has to fit on one card, so with heavy textures, dense scans or large .blend files, GPU memory decides whether the render is possible.
GPUs and GPU memory compared
Figures are each company's published specs, linked as the source. Services whose official pages do not name the GPU are not listed.
| Service | GPU for Cycles | GPU memory per card | Published price unit (lowest listed) |
|---|---|---|---|
| JANCTION Render | NVIDIA RTX PRO 6000 Blackwell, every job | 96 GB | 0.1 JPY per GPU-second (360 JPY per GPU-hour) |
| GarageFarm.NET | RTX 6000 Pro Blackwell 96GB; also RTX 4000 Ada, L40s, RTX A5000 | 96 GB on Medium and High; up to 48 GB on Low | from $0.004 per OB-hour (Low) |
| BlendSwap Render Farm | NVIDIA L40S | 48 GB per worker | $4.30 per server-hour |
| RenderStreet | NVIDIA L40s | 48 GB per server | $4.49 per server-hour |
| Renderjuice | primarily RTX 4090s and 5090s | "up to 48 GB" (RTX 4090: 24 GB, RTX 5090: 32 GB per card) | from $49 a month for 2,000 GPU minutes |
| Ranch Computing | GeForce RTX 2080 Ti, 3090, 4090, 5090 | 11 to 32 GB | from 0.005 EUR per OB-hour (excl. tax) |
| iRender (machine rental) | RTX 4090 or RTX 5090 servers | 24 or 32 GB (a larger scene uses system RAM, slower) | from $8.20 per hour |
| RebusFarm | Quadro RTX 6000 | not stated | 0.45 cent per OB-hour |
| SheepIt (volunteer farm) | volunteers' own GPUs (reference RTX 3060) | depends on the volunteer | free |
OB-hour is the farms' own benchmark-based unit, so those prices do not convert directly to GPU-hours. Specs can change; follow the links for the current values.
GPU memory decides what renders
Cycles keeps the scene's textures, geometry and acceleration structures on the GPU while it renders, and one frame runs on one card. A scene that needs 40 GB falls back to slower system memory or stops with an out-of-memory error on a 24 GB card. The scenes that hit this first are the ones people send to a farm: production .blend files with packed 8K textures, photogrammetry and Gaussian-splat scans with millions of points, and dense CAD or terrain meshes. On a 96 GB card they fit without shrinking textures or splitting the scene.
Running Blender on a GPU cloud yourself
General GPU clouds rent the same class of card. On their pricing pages (November 16, 2026), RunPod lists the RTX PRO 6000 96GB from $1.69 an hour (Community Cloud) or $2.69 (Secure Cloud), and Modal at $0.000842 a second (about $3.03 an hour). That can be cheaper per hour, but you set up Blender and the job splitting yourself and pay while the machine starts, waits and idles. JANCTION Render charges only the GPU-seconds a render uses, AI agents call it directly over MCP, and monthly plans keep renders and scenes in JANCTION Storage (beta) for as long as you are subscribed.
How we compared
On November 16, 2026 we read each company's pricing, specification, FAQ and documentation pages. "GPU memory per card" is the published value for one card. The comparison is also on render.janction.jp and is reviewed every month.
Frequently asked questions
- Which render farms give a Blender job 96 GB of GPU memory?
- On their own pages (checked November 16, 2026), GarageFarm offers a 96 GB RTX 6000 Pro Blackwell on its Medium and High tiers, and JANCTION Render runs every job on a 96 GB RTX PRO 6000 Blackwell. Most other farms list 24 to 48 GB cards.
- Does a bigger GPU make Cycles faster?
- Not by itself. GPU memory decides whether a scene fits; speed depends on the card and the scene. A scene that does not fit falls back to slower system memory or stops, so on heavy scenes memory decides whether the render is possible.
- Is renting an RTX PRO 6000 on a GPU cloud cheaper?
- Per hour it can be (RunPod from $1.69 an hour, Modal about $3.03, November 16, 2026). You then set up Blender and the job splitting yourself and pay while the machine starts and idles; JANCTION Render charges only the GPU-seconds a render uses.
About JANCTION Render
JANCTION Render is JasmyLab's Blender cloud GPU rendering service for AI agents. Add one MCP endpoint to ChatGPT, Claude, Codex or Cursor, or send a file from the browser, and your AI can render Blender images and videos without a local GPU or Blender.
- Service: https://render.janction.jp/
- MCP endpoint: https://render.janction.jp/mcp
- GPU and VRAM comparison on the service site: https://render.janction.jp/blender-render-farm-gpu-memory
