The cheapest GPU for Blender renders isn't the fastest one — RTX A4000 in a cost-per-frame benchmark
240 frames of a car turntable at 1080p, Cycles OptiX 128 samples: 3 min 18 s, US$0.014 total. The RTX A4000 renders slower than an H100 — but each frame costs a tenth as much. Why ‘fastest GPU’ is the wrong metric in the cloud.
Every 3D studio founder who asks us about cloud rendering has the same reflex: "I'll take the strongest card." H100. Or at least a 4090. Anything weaker sounds like compromise.
That's a very expensive metric. Cloud rendering isn't paid per GPU-hour — it's paid per finished frame. And the card that produces the cheapest frame on real Blender workloads isn't the fastest one, it's usually the second-cheapest.
The test
We ran a standard Blender scene across three GPUs: a 240-frame Audi RS6 turntable, Cycles OptiX, 128 samples per frame, 1920 × 1080, AgX color transform. Not marketing cinema — the sort of thing a motion-graphics studio ships to a client Monday morning.
- RTX A4000 — 3 min 18 s, US$0.014 total, 0.006 ¢/frame
- RTX 4090 — 1 min 32 s, US$0.028 total, 0.012 ¢/frame
- H100 (80 GB) — 41 s, US$0.081 total, 0.034 ¢/frame
The H100 is 4.8× faster than the A4000 — and 5.8× more expensive per frame. On a wall-clock metric it wins. On an invoice metric it loses badly.
Why the ratio behaves this way
Cloud GPU prices don't scale linearly with rendering throughput. Two effects:
-
RAM is the price driver, not compute. An H100 has 80 GB HBM3 on a 3 TB/s bus. For a Blender scene that fits in 3 GB of VRAM, that's overkill — the price subsidises workloads we don't care about (LLM inference, scientific simulation). The A4000 has 16 GB GDDR6 on 448 GB/s — matches 95 % of commercial Blender scenes like a hand in a glove.
-
OptiX turns ray-tracing into hardware primitives. Cycles is OptiX-accelerated on every RTX card. The bandwidth gap between A4000 and H100 shows up in the denoising kernel at the same factor — not the 8× GPU-price gap.
The net: choosing the A4000 costs you 2 minutes on a 4-minute job, but wins a factor of 5 on the invoice. For a small studio owner who doesn't have a cloud budget on autopilot, that's not nice-to-have — that's the difference between "cloud pays off" and "I'll stick with my old MacBook".
When the faster GPU is still right
Edge cases exist. An 8K environment render with a 32 GB VRAM peak will OOM on the A4000 halfway through — then you need the 4090 (24 GB) or A6000 (48 GB). A client wants preview in 20 minutes instead of 3 hours — the fast turnaround pays for the extra compute. But these are exceptions, not the default.
The default should always be the cheapest GPU that can render the scene without a VRAM OOM. For any 1080p Cycles standard job under 12 GB VRAM, that's the A4000. For 4K or character animations with sub-D meshes: 4090. Above that: A6000. We practically never see H100s used on commercial Blender jobs — it's a tool for LLM fine-tuning that only appears in the catalog because someone might want it.
The meta-lesson
Every "strongest card" recommendation is a restaurant wine sale — the margin is in the upsell, not the base product. Cloud GPU vendors like AWS/GCP have every incentive to advertise the enterprise cards. We're the opposite: the cheaper your frame, the more often you come back.
The right decision score is not min/frame. It's €/frame with a 24-hour deadline buffer. The A4000 wins that score in almost every Blender-studio everyday scenario that isn't an exception.
Reproduce it
All three benchmarks ran on ZER0ONE Lab Cinema Render with identical Cycles parameters and the same Blender 4.2. The A4000 is our default tier — you can rebuild the turntable job with one click and have your own cost-per-frame line on your invoice within five minutes.
If you're not sure which GPU fits your scene: send us the .blend, we tell you which card makes sense and what the job will cost before you even sign up for the beta. Under 24 h reply.