TL;DR:
It costs about $30–50K SGD to build a dedicated, rack-mounted GPU render farm with around 10 render nodes across 3 machines. But don’t forget: budget for power headroom and heat exhaust.

And yes — in my opinion, it’s absolutely worth it.

Backstory

Before GPU Rendering Was a Thing

When we started the studio, we ran with two used Lenovo workstations bought from theserverstore. Back then, CPU rendering was the only option so we needed an affordable solution for high core-count cpus. New dual xeon workstations were WAY too expensive running around 10-20K a piece. So buying a used computer from the USA at $1600 SGD a piece made extreme financial prudent sense. These were no slouch of a machine, with each giving us 24 threads total compared to the usual 8 for an i7 computer.

Our very first workstations were used Lenovo D30 workstations with 24 Threads and 128 GB of ram. Super heavy and durable! We used them until 2022!
Back then, you could swarm your entire network into rendering a single frame — one massive CPU render per image. It was magical.

The Shift to GPU Rendering

Not too long after 2017, GPU rendering became the place to be if you wanted to have the edge in speed and quality. With this, the need for high core count CPUs faded. This was due to the fact that we could render frames that would have taken 20-40 minutes in only 3-5 minutes.

Case in point:
Our first fully GPU-rendered project, the Razer Basilisk Ultimate Commercial, averaged 2.5 minutes per frame. The entire production wrapped in under five weeks.

The Rise of Affordable Power: Threadrippers & 1080Tis

After the 2018 crypto crash, we scooped up many GTX 1080Tis for ~$700 SGD each, making the switch to GPU rendering incredibly cost-effective.

But we quickly hit the PCIe lane bottleneck. Consumer CPUs couldn’t handle multi-GPU loads efficiently — which is where Threadripper came in. With enough lanes for 3–4 GPUs per machine, and great performance-to-cost, it let us build the next generation of artist workstations.

Do You Still Need Multi-GPU Setups?

Honestly, for regular artist workstations, not really — modern GPUs are beasts.
At one point, the only way you could get meaningful power and speed from the GPU render was to have multiple cards stacked up in a single machine. Also back then Nvidia themselves were pushing the link technology for their cards; Right now they don’t even support it and are instead pushing out single cards that are more powerful and have larger VRAM.

This means whilst single cards are much more expensive, they are also kind of like multiple cards in one.

For example:

  • 3× 1080Ti (10K CUDA cores, 11GB VRAM per card)
    = 1× RTX 3080Ti (10K CUDA cores, 12GB VRAM per card)
    = 1× RTX 5090 (21K CUDA cores 32GB VRAM per card)

I think you get the point. There actually is alot of benefit to running single cards for regular workstations.

So instead of stacking six cards like before, we now give each artist 1–2 high-end cards, reducing heat, noise, and power draw.

Things to Watch Out For

Power Draw, Heat Dissipation, and Cost are three important factors to consider.

1. Power Draw

This might be the most important consideration of all — because regardless of your budget or rendering needs, physics doesn’t care. If your render farm’s total power draw isn’t planned and provisioned for properly, you will run into:

  • Short circuits during peak loads
  • Random, unrecoverable shutdowns
  • And worst of all — a real risk of electrical fires

No amount of expensive GPUs or optimized software matters if your breakers can’t handle the load.

Modern GPUs like the RTX 5090 pull 575W each.
3 of those? That’s ~1.8kW per machine, which is dangerously close to the 13A/3,000W limit per power socket.

We upgraded our MCBs to 32A breakers to support two machines per circuit safely. Stability matters more than raw speed — a short-circuit during a render crunch is a nightmare.

2. Heat & Noise

Packing multiple hot cards into one box heats up your studio, and can lead to a noisy, stressful workspace.

We solved this by:

  • Keeping artist workstations in the work area leaner (1–2 GPUs)
  • Moving dense GPU machines to a dedicated server room with ventilation.
    • It’s super important to install some form of a ventilation system that extracts the heat out of the server room if you don’t want everything to overheat and melt.
Server exhaust systems could get very expensive, but what we did was to get some guys who do extractions for industrial kitchens, and they installed a simple vent for us for a relaatively low cost of $1.2K SGD.

3. Cost

GPUs are now marked up and cost anywhere from $2.5K–6K SGD due to excessive AI demand.
So we’ve chosen to stick with 30-series GPUs like the 3080Ti. And if we decide to upgrade in the future we can pool existing 30-series cards together, and install single new generation cards giving us a hybrid composition of 2*30-series machines and 1*50-series machines.

Cost Breakdown:

  • Hardware (3 machines, 10 GPUs RTX 3080Ti): ~$36K SGD (incl. the exhaust system)

Do you save money with running a local render farm?

Well let’s look at it from a 3 year running point of view. Our local render farm tackles about 90% of all our work, and we use Cloud rendering for the remaining 10%, which is mostly during crunch time when our farm just can’t handle the bottleneck.

Local Render farm setup: $~36K SGD
Cloud rendering (10% overflow work): ~$2K/year
Total cost over 3 years = $42K SGD

If we went 100% cloud:
$2K = 10% → ~$20K/year → $60K over 3 years

Savings: ~$18K SGD, plus full creative control.

This cost savings might not seem like a huge difference in the grand scheme of a business, but it adds up — especially when you factor in profit margins, infrastructure upgrades, welfare, and bonuses. All that ties back into what makes a healthy studio function, and every dollar counts.

Our Current Setup: 10 Render Nodes, 3 Machines

Not counting the artists’ workstations that double up as a render node when not in use, as most of the time thoses machines can only be used after work hours, and the idle artist waiting for renders is not a great use of their time.

Here’s our in-house render farm:

  • 2× Threadripper PRO servers, each with 4× RTX 3080Ti
  • 1× Dual-Xeon Supermicro server with 2× RTX 3080Ti (from 2017)

Key Insight:
Some render licenses like Redshift are charged per machine, not per GPU.
So 4 render nodes in 1 machine = 1 license = massive savings.

These three machines take up just 6U of rack space and handle 90% of our rendering.

With only 3 Machines, taking up 12Us on our 42U rack. And these run 10 GPU render nodes utilising only 3 redshift licenses.

So, Is It Worth It?

Absolutely.
But not just for cost savings.

For me, I always found the joy in enabling the people around me as much as I can. Sometimes as a director, you don’t get to run with everyone’s ideas, due to time, constraints or just for the fact that sometimes you just can’t. So it really does bring me joy to watch the infrastructure I’ve built for the studio enable our artists to play and explore without compromise or cost.

Over the years, I’ve enjoyed watching our artists utilise studio resources in personal work when there’s some down time; The fact that the idle farm is around and ready for them to use whenever they want and can is something that really makes me happy to see.

Below you can see some of the beautiful animated works they’ve done in the studio and rendered on our in-house farm.

Super cute little doodle by our Lead Lighting CGI Artist, Gina See.
Full-CGI waterfall and lake simulation and render by our TD and FX Artist, Alex Seow.
Adorably Tactile retro-printer doodle by our CGI Generalist, Loy Qian Ling.
On-going stylised 3D animated character work by one of our Lead CGI Artists, Benjamin Lim.

The Intangible Benefit: Creative Freedom

As much as I’m. big fan of rendering on the cloud, it always feels like spending. And with spending, there comes hesitation. And that hesitation — even when small — compounds.

When creatives hesitate, they explore less.

A local render farm? It’s just there. Paid for. Idle. Waiting.
And when artists can play freely, without fear, creativity flows.

This isn’t limited to just personal creative work like I’ve shared, but also on the copious amount of iterations of renders for the work we produce for our clients. I can’t share them but it’s definitely ALOT. Terabytes.


Final Thoughts

Should you build a local render farm?

  • If you have someone who’s technically inclined, willing to maintain the system, YES.
  • If not, or you don’t have time, Go cloud.

For us at Masonry, the hybrid system gives us:

  • Creative control
  • Cost efficiency
  • Faster iteration
  • A more playful production culture

So yes, to me — it’s worth it.
And it’s one more piece of the infrastructure puzzle that helps us keep building really amazing image work that lasts.

The first iteration of Masonry Studios’s setup.
Nicholas Chia

Author Nicholas Chia

As a Creative Director, I craft content that blends creativity with rhythmic execution. With expertise in CGI, video, and audio design, I’ve shaped Masonry Studios through innovation in creative, business, and production. Exploring the intersection of art, technology, and strategy, I write and share insights that shape our studio's work.

More posts by Nicholas Chia