SuperRenders Render Farm

09/21/2026

GPU gets the hype. CPU still does 70% of the actual rendering work on our farm.

Why: CPU nodes carry far more system RAM (96-256GB vs 16-32GB of GPU VRAM), and CPU engines still lead on volumetrics, hair, and subsurface scattering.

GPU is growing fast. It hasn't replaced CPU - the job log says otherwise.

Full guide → https://superrendersfarm.com/article/what-is-a-render-farm-guide

09/21/2026

Most people use "render farm" and "render service" as the same word. They're not.

Render farm = the hardware running your job. Render service = the business wrapped around it - who installs the software, handles licensing, and catches a failed node before you notice.

We run fully managed: upload, render, download. 20,000+ CPU cores, dedicated RTX 5090 machines, licensing bundled into the rate.

Side-by-side breakdown → https://superrendersfarm.com/article/render-service-vs-render-farm

09/21/2026

AWS Deadline Cloud has a documented gap, public since launch: the MayaUSD plugin isn't in the Maya Conda package it ships (GitHub issue #409). For a Maya + USD + Arnold pipeline, that's scenes failing, not a footnote.

Deadline Cloud orchestrates well. It doesn't install, mount, license, or match your environment. That's still on you.

Full breakdown → https://superrendersfarm.com/article/aws-deadline-cloud-alternative

09/20/2026

37.6 GB logged on a card with 32 GB onboard. That's what one production Cycles job hit on our RTX 5090 node, more than the card's own onboard memory. Most jobs stay light (median 5.6 GB, 90th percentile 11.5 GB across 57 logged jobs, 51 days of production data), but we size VRAM for the tail, not the median. Full numbers:

The full data behind our CGW SIGGRAPH field report: RTX 5090 speedup, VRAM percentiles, denoise rate, and driver stability from 7 weeks of production.

08/13/2026

XGen fur vanishing on a render farm looks like a software bug. It almost never is.

XGen doesn't store groom or fur data inside the .ma/.mb scene file. It references external folders instead. Miss one of them in the handoff and the render finishes clean, no error, no warning, just no fur.

Full breakdown, incl. the same failure class with Arnold's .tx cache:

Why XGen grooms and .tx textures fail on render farms - unmapped paths, stale caches, maketx mismatches - plus a preflight checklist that prevents it.

08/12/2026

About 90 people a month search "Pixar render farm." Pixar has never rented out a single node.

What they're actually after is RenderMan, Pixar's licensed renderer. It runs on our farm too, alongside Arnold, V-Ray, and Redshift.

The catch was always licensing, not availability. Full breakdown:

Searching for a 'Pixar render farm'? Here's what that means in practice: RenderMan for Maya on a managed cloud farm, licensing and setup explained.

08/11/2026

Apple Silicon is now in our cost calculator - M1 through M5, twenty-six chips.

One number to know before you use it: if your workflow is Redshift Metal, Octane X, or Blender Metal, actual render time runs 2-4× faster than our CPU-path estimate. We surface that caveat inline so the number you're reading doesn't mislead.

What else changed in the May 2026 update ->

Apple Silicon, current-gen Intel/AMD, full GPU dropdown, OctaneBench 2025.2.1, and saved settings — what's new in our cost calculator this May.

08/10/2026

150 hours of GPU rendering per month. That's the break-even for a solo RTX 5090 workstation in 2026 - before you factor in the number most freelancer spreadsheets skip entirely.

When your only machine is rendering, it isn't available for client revisions.

For anyone below 150 hours/month, cloud wins on direct cost alone. The full math - including hybrid model - is here: https://superrendersfarm.com/article/single-5090-workstation-vs-cloud-rendering-cost-2026

08/07/2026

Two GPUs in one rental node doesn't mean 64GB of VRAM to work with.

Each RTX 5090 card carries 32GB. The two pools don't combine - 32GB per card is the real ceiling any single frame has to fit inside. A second card buys you a second frame rendering in parallel, not more headroom for one.

Full FAQ breakdown ->

How dedicated GPU node rental compares to per-frame cloud rendering: what you get, how billing works, and when a dedicated node makes sense.

08/07/2026

Most GPU benchmarks do not tell you if they are trustworthy. Ours has to pass five checks first:

One render task per card. Matched pairs only. One driver for the whole test window. Same denoiser on both sides. Warm run, not a cold start.

Skip any one of these and the number is measuring the setup, not the hardware.

Full checklist -> https://superrendersfarm.com/article/render-farm-gpu-benchmark-methodology-2026