# How Superluminal is so fast

> Superluminal posted the shortest render pipeline and delivered complete local output first in all four benchmark scenes. The margin comes from removing work around each frame.

- Canonical HTML: https://superlumin.al/how-superluminal-is-so-fast
- Markdown endpoint: https://superlumin.al/how-superluminal-is-so-fast.md
- Structured page data: https://superlumin.al/how-superluminal-is-so-fast.json
- Published: 2026-08-05
- Last updated: 2026-08-06
- Author: Superluminal Engineering

## Quick answer

Benchmark results show how Superluminal removes delays through its Persistence Engine, concurrent output work, streaming delivery, and farm scheduler.

### Key benchmark facts

| Fact | Value | Context | Evidence |
| --- | --- | --- | --- |
| Fastest render pipelines | Four of four benchmark scenes | Superluminal recorded the shortest render pipeline in all four benchmark scenes. |  |
| Classroom scaling | 411.0 s to 90.5 s | Farm time after upload at 5 and 30 selected nodes. |  |
| Blender reuse | One Blender process for later frames | Later frames in the same job can reuse the process. |  |
| Output delivery | Finished outputs transfer during rendering | The add-on retrieves available files while later frames continue. |  |

## Data access

- [This page as structured JSON](https://superlumin.al/how-superluminal-is-so-fast.json)

- Every chart below includes a full-resolution PNG and its underlying values as a Markdown table.

## Related documentation

- [Automation and public API](https://superlumin.al/docs/automation): Use Superluminal's experimental public API and agent skills with explicit safeguards around billable render submission.
- [Render-node and experimental settings](https://superlumin.al/docs/blender-addon/render-node-settings): Choose the farm Blender version and understand Finish Frame On Error and the Persistence Engine.
- [Choose an upload mode](https://superlumin.al/docs/blender-addon/upload-settings): Choose ZIP or PROJECT packaging and configure project-root handling for reliable, repeatable submissions.
- [Core concepts](https://superlumin.al/docs/concepts): Understand how organizations, projects, jobs, frames, and credits fit together in Superluminal.
- [Web app](https://superlumin.al/docs/web-ui): Navigate Superluminal projects, render jobs, detailed frame review, billing, account settings, Market, and support.

## Complete report

Superluminal delivered all four benchmark scenes to local disk first and recorded the shortest [render pipeline](/compare/blender-render-farms) in the three Cycles scenes. On Classroom, 30 nodes completed the pipeline in 90.5 seconds. The next fastest tested service took 163.6 seconds.

That 73.1-second lead came from the complete system, not a faster Blender setting. At Superluminal, speed is a product requirement, so we measure and remove delays in submission, scheduling, process startup, output handling, and delivery. Most of that work is explained below. One part is not.

![Blender render farms compared by recorded render-pipeline time and cost in 43 recorded runs](https://superlumin.al/seo/charts/best-blender-render-farms-time-vs-cost-2026.png "The benchmark result behind the speed work")

*[Open the full-resolution PNG](https://superlumin.al/seo/charts/best-blender-render-farms-time-vs-cost-2026.png).*

#### Underlying data: The benchmark result behind the speed work

X axis: **Render pipeline, seconds**. Y axis: **Cost, USD**.

| Panel | Render farm | Run | X | Y | Marker | Status | Recorded detail |
| --- | --- | --- | --- | --- | --- | --- | --- |
| All runs | Blendergrid | Blendergrid | 5123.701 | 11.73 | circle | completed | classroom: 5123.7 s render pipeline, $11.73 (displayed quote, not a final debit) |
| All runs | Blendergrid | Blendergrid | 4452.934 | 7.3 | square | completed | junkshop: 4452.9 s render pipeline, $7.30 (displayed quote, not a final debit) |
| All runs | Blendergrid | Blendergrid | 3747.635 | 4.1 | diamond | completed | monster: 3747.6 s render pipeline, $4.10 (displayed quote, not a final debit) |
| All runs | Fox Renderfarm | Fox Renderfarm | 292.833 | 3.204 | circle | completed | classroom: 292.8 s render pipeline, $3.20 (exact charge) |
| All runs | Fox Renderfarm | Fox Renderfarm | 540.434 | 2.443 | triangle | completed | fox-eevee: 540.4 s render pipeline, $2.44 (exact charge) |
| All runs | Fox Renderfarm | Fox Renderfarm | 710.566 | 2.456 | square | completed | junkshop: 710.6 s render pipeline, $2.46 (exact charge) |
| All runs | Fox Renderfarm | Fox Renderfarm | 575.033 | 1.795 | diamond | completed | monster: 575.0 s render pipeline, $1.79 (exact charge) |
| All runs | GarageFarm | GarageFarm | 163.6 | 7.55 | circle | completed | classroom: 163.6 s render pipeline, $7.55 (exact charge) |
| All runs | GarageFarm | GarageFarm | 113.966 | 1.12 | triangle | completed | fox-eevee: 114.0 s render pipeline, $1.12 (exact charge) |
| All runs | GarageFarm | GarageFarm | 127.333 | 3.92 | square | completed | junkshop: 127.3 s render pipeline, $3.92 (exact charge) |
| All runs | GarageFarm | GarageFarm | 88.566 | 2.73 | diamond | completed | monster: 88.6 s render pipeline, $2.73 (exact charge) |
| All runs | iRender | iRender | 1225.7 | 10.463 | circle | completed | classroom: 1225.7 s paid workstation window, $10.46 (rate-derived upper bound) |
| All runs | iRender | iRender | 736.733 | 6.139 | triangle | completed | fox-eevee: 736.7 s paid workstation window, $6.14 (rate-derived full upload-through-local-completion machine cost, 6.1394 iR at the verified 30 iR per hour rate, converted at the $1 face value per main-balance iR actually deducted; a 50 iR promotional balance from the same deposit could halve the effective rate but its spendability is unconfirmed) |
| All runs | iRender | iRender | 1033 | 8.608 | square | completed | junkshop: 1033.0 s paid workstation window, $8.61 (rate-derived full upload-through-local-completion machine cost, 8.6083 iR at the verified 30 iR per hour rate, converted at the $1 face value per main-balance iR actually deducted; a 50 iR promotional balance from the same deposit could halve the effective rate but its spendability is unconfirmed) |
| All runs | iRender | iRender | 617.1 | 5.143 | diamond | completed | monster: 617.1 s paid workstation window, $5.14 (rate-derived full upload-through-local-completion machine cost, 5.1425 iR at the verified 30 iR per hour rate, converted at the $1 face value per main-balance iR actually deducted; a 50 iR promotional balance from the same deposit could halve the effective rate but its spendability is unconfirmed) |
| All runs | RebusFarm | RebusFarm | 445.533 | 11.45 | circle | completed | classroom: 445.5 s render pipeline, $11.45 (dated or provider-defined USD estimate) |
| All runs | RebusFarm | RebusFarm | 168.367 | 15.18 | triangle | completed | fox-eevee: 168.4 s render pipeline, $15.18 (dated or provider-defined USD estimate) |
| All runs | RebusFarm | RebusFarm | 315.5 | 10.34 | square | completed | junkshop: 315.5 s render pipeline, $10.34 (dated or provider-defined USD estimate) |
| All runs | RebusFarm | RebusFarm | 101.133 | 3.63 | diamond | completed | monster: 101.1 s render pipeline, $3.63 (dated or provider-defined USD estimate) |
| All runs | Render Network | Render Network | 1186.565 | 14.65 | circle | completed | classroom: 1186.6 s render pipeline, $14.65 (dated or provider-defined USD estimate) |
| All runs | RenderJuice | RenderJuice | 289.333 | 1.57 | circle | completed | classroom: 289.3 s render pipeline, $1.57 (dated or provider-defined USD estimate) |
| All runs | RenderJuice | RenderJuice | 342.301 | 0.83 | triangle | completed | fox-eevee: 342.3 s render pipeline, $0.83 (dated or provider-defined USD estimate) |
| All runs | RenderJuice | RenderJuice | 234.067 | 0.98 | square | completed | junkshop: 234.1 s render pipeline, $0.98 (dated or provider-defined USD estimate) |
| All runs | RenderJuice | RenderJuice | 204.167 | 0.91 | diamond | completed | monster: 204.2 s render pipeline, $0.91 (dated or provider-defined USD estimate) |
| All runs | RenderStreet | RenderStreet | 530.3 | 4.37 | circle | completed | classroom: 530.3 s render pipeline, $4.37 (exact charge) |
| All runs | RenderStreet | RenderStreet | 813.4 | 0.83 | triangle | completed | fox-eevee: 813.4 s render pipeline, $0.83 (exact charge) |
| All runs | RenderStreet | RenderStreet | 345.667 | 2.91 | square | completed | junkshop: 345.7 s render pipeline, $2.91 (exact charge) |
| All runs | RenderStreet | RenderStreet | 373.567 | 2.12 | diamond | completed | monster: 373.6 s render pipeline, $2.12 (exact charge) |
| All runs | Superluminal | 10 GPUs | 226.733 | 1.35 | circle | completed | classroom: 226.7 s render pipeline, $1.35 (exact charge) |
| All runs | Superluminal | 30 GPUs | 90.467 | 1.9 | circle | completed | classroom: 90.5 s render pipeline, $1.90 (exact charge) |
| All runs | Superluminal | 5 GPUs | 411 | 1.1 | circle | completed | classroom: 411.0 s render pipeline, $1.10 (exact charge) |
| All runs | Superluminal | 10 GPUs | 66.467 | 0.33 | triangle | completed | fox-eevee: 66.5 s render pipeline, $0.33 (exact charge) |
| All runs | Superluminal | 30 GPUs | 75.567 | 0.31 | triangle | completed | fox-eevee: 75.6 s render pipeline, $0.31 (exact charge) |
| All runs | Superluminal | 5 GPUs | 77.733 | 0.14 | triangle | completed | fox-eevee: 77.7 s render pipeline, $0.14 (exact charge) |
| All runs | Superluminal | 10 GPUs | 125.067 | 0.64 | square | completed | junkshop: 125.1 s render pipeline, $0.64 (exact charge) |
| All runs | Superluminal | 30 GPUs | 64.5 | 1.21 | square | completed | junkshop: 64.5 s render pipeline, $1.21 (exact charge) |
| All runs | Superluminal | 5 GPUs | 200.734 | 0.48 | square | completed | junkshop: 200.7 s render pipeline, $0.48 (exact charge) |
| All runs | Superluminal | 10 GPUs | 112.867 | 0.62 | diamond | completed | monster: 112.9 s render pipeline, $0.62 (exact charge) |
| All runs | Superluminal | 30 GPUs | 59.933 | 0.98 | diamond | completed | monster: 59.9 s render pipeline, $0.98 (exact charge) |
| All runs | Superluminal | 5 GPUs | 197.667 | 0.5 | diamond | completed | monster: 197.7 s render pipeline, $0.50 (exact charge) |

## More nodes solve only part of the wait

Classroom's render pipeline fell from 411.0 seconds on five nodes to 90.5 seconds on 30 nodes. Six times as many nodes produced a 4.5-fold gain.

The result is useful, but it also exposes the fixed-cost floor. Every job still starts workers and Blender, opens the project, and waits for its last group of frames. More capacity cannot divide those costs by six.

## The Persistence Engine pays startup once

Starting Blender and opening the same project for every frame repeats work that cannot improve the image. The Persistence Engine keeps one Blender process available for later frames from the same job.

After a frame finishes, the engine changes the frame number and output path before it starts the next render. A new job or a recovery starts a new Blender process, which keeps reuse inside the job that created it.

![Fresh Blender processes compared with the Persistence Engine across three frames](https://superlumin.al/seo/charts/persistence-engine-process-reuse-2026.png "One Blender launch serves many frames")

*A new job or recovery starts a new Blender process. [Open the full-resolution PNG](https://superlumin.al/seo/charts/persistence-engine-process-reuse-2026.png).*

#### Underlying data: One Blender launch serves many frames

| Mode | Blender launches | Project opens | Frames |
| --- | --- | --- | --- |
| Fresh process per frame | 3 | 3 | 3 |
| Persistence Engine | 1 | 1 | 3 |
| New job or recovery | Starts a new process | Opens the project | Continues the new process |

## A controlled test isolates the saved time

A controlled Blender 5.1.1 test rendered the same six-frame Cycles animation in two ways. Six fresh processes took a median 2.943 seconds, while the Persistence Engine took 0.849 seconds with one process.

The Persistence Engine removed 2.094 seconds, or 71.2 percent of the measured wall time. All 42 paired frame comparisons had identical decoded pixels.

![Fresh Blender processes took 2.943 seconds and the Persistence Engine took 0.849 seconds](https://superlumin.al/seo/charts/blender-5-1-process-reuse-wall-time-2026.png "One Blender launch cut wall time by 71.2 percent")

*Median local wall time across seven paired repeats. All paired frame outputs had identical decoded pixels. [Open the full-resolution PNG](https://superlumin.al/seo/charts/blender-5-1-process-reuse-wall-time-2026.png).*

#### Underlying data: One Blender launch cut wall time by 71.2 percent

| Mode | Median wall time | Blender launches |
| --- | --- | --- |
| Fresh process per frame | 2.943319 s | 6 |
| Persistence Engine | 0.848870 s | 1 |

[Open the controlled benchmark data](https://superlumin.al/seo/evidence/blender-process-reuse-benchmark-2026-08-05.json)

This test used one Apple M5 Pro, CPU rendering, one small synthetic scene, and seven paired repeats. It excluded the network, queue, scheduler, upload, and download path.

The test therefore measures the principle, not production farm speed. Long frames make startup a smaller percentage of total time, but every reused frame still avoids another Blender launch and project open.

The [render node settings guide](/docs/blender-addon/render-node-settings#persistence-engine) explains when to enable the Persistence Engine and how to check scene compatibility.

## Render nodes should spend their time rendering

A finished frame still needs output preparation and upload before you can use it. If the render node performs that work in sequence, the GPU can sit idle before it receives another frame.

Superluminal moves finished output to separate background workers. The render node can request its next frame while those workers prepare and upload the earlier result.

This overlap gives you two benefits. Paid render capacity returns to Blender sooner, and completed files can start moving toward your workstation before the job ends.

## Download frames while later frames render

The Blender add-on checks for new output throughout the job and downloads available files in parallel. You can inspect, edit, or composite early frames without waiting for one final archive.

![Render nodes, output workers, storage, and the Blender add-on working on finished frames at the same time](https://superlumin.al/seo/charts/async-render-output-download-flow-2026.png "A finished frame can reach local disk before the job ends")

*Rendering, output preparation, upload, and local download can overlap. [Open the full-resolution PNG](https://superlumin.al/seo/charts/async-render-output-download-flow-2026.png).*

#### Underlying data: A finished frame can reach local disk before the job ends

| Stage | Can overlap with |
| --- | --- |
| Render a later frame | Prepare and upload an earlier frame |
| Prepare and upload output | Render work on other frames |
| Download available files | Later render and upload work |

[Open recorded run events](https://superlumin.al/seo/evidence/render-farm-benchmark-run-events-2026-08.json)

The [download guide](/docs/blender-addon/manage-and-download#download-output) covers live polling, resume behavior, and original output files. [Live previews](/live-previews) explains browser review.

## The scheduler is the part we keep secret

Every mechanism on this page has a public explanation. The scheduler does not, and that is deliberate. Its placement rules are the proprietary core of the farm. After every status change, they decide which ready frame lands on which node. Nothing waits for a batch boundary, and nothing idles while a decision is pending.

Frame ordering, node selection, and failure recovery run beside those rules as one tuned system. A slow decision between frames wastes the same user time as a slow render, so we engineer decision time down the way other farms chase render time.

You can see what the rules do without seeing them. [Ten Superluminal nodes](/compare/blender-render-farms#node-count) finished Classroom's render pipeline ahead of a farm that showed 129 nodes running. That gap is not hardware. It is placement.

The [full benchmark](/compare/blender-render-farms) publishes every recorded run, charge, and timing boundary. The placement rules stay ours.

## Sources

| # | Source | Type | Checked | URL |
| --- | --- | --- | --- | --- |
| 1 | Render farm benchmark dataset, July to August 2026 | First-party benchmark evidence | 2026-08-05 | https://superlumin.al/seo/evidence/render-farm-benchmarks-2026-08.json |
| 2 | Render farm benchmark run events | First-party timing and interface evidence | 2026-08-05 | https://superlumin.al/seo/evidence/render-farm-benchmark-run-events-2026-08.json |
| 3 | Controlled Blender 5.1 process-reuse benchmark | First-party controlled benchmark evidence | 2026-08-05 | https://superlumin.al/seo/evidence/blender-process-reuse-benchmark-2026-08-05.json |
| 4 | Superluminal Blender add-on documentation | Product documentation | 2026-08-05 | https://superlumin.al/docs/blender-addon |