GPU rendering servers for Blender, Redshift and Octane

GPU rendering servers make sense when a deadline needs more frames than your workstations can deliver. We rent NVIDIA RTX PRO 6000 Blackwell and L40S GPUs by the month, as single-GPU servers or as one 8-GPU server with 768 GB of total VRAM, for Blender Cycles, Redshift, Octane and other GPU renderers. You get root access over SSH, local NVMe storage and hardware that no other customer shares.

Which GPUs are built for rendering

Path tracers such as Cycles with OptiX, Redshift and Octane use the RT cores of NVIDIA's RTX-class GPUs to accelerate ray traversal. The RTX PRO 6000 Blackwell, the L40S and the RTX 5090 have them. The H100, H200 and B200 are AI accelerators without RT cores, so they are the wrong rental for path tracing, however fast they are at training.

• RTX PRO 6000 Blackwell 96 GB of GDDR7 with ECC, 1.6 TB/s, PCIe Gen5
• L40S 48 GB of GDDR6, 864 GB/s, PCIe Gen4
• 8× RTX PRO 6000 server 768 GB of total VRAM, 2× EPYC 9555 with 128 cores, 1.5 TB of DDR5

Before booking, check that your renderer and plugin versions support Blackwell GPUs, since older releases may not recognize them. Our RTX PRO 6000 vs L40S comparison sets the two render cards side by side.

How much VRAM heavy scenes need

Most GPU renderers need the scene, its textures and the acceleration structure in each GPU's memory. Some can page data to system RAM when a scene does not fit, at a cost in speed; others stop with an out-of-memory error. Texture math adds up quickly: one 8K texture (8192 × 8192, RGBA) is about 268 MB uncompressed at 8 bits per channel and 537 MB at 16 bits, so a hundred 8K maps can take about 27 GB before geometry, hair, volumes and the BVH.

• 48 GB on the L40S: product shots, architectural visualization, motion graphics and lighter VFX shots.
• 96 GB on the RTX PRO 6000: dense environments, high-resolution volumes, heavy hair and fur, and shots that fail on 24–48 GB cards.

Setting up GPU rendering servers for a production

These are headless Linux servers reached over SSH, which suits command-line rendering, render managers such as Flamenco or your studio's own, and look-dev driven remotely. If your workflow depends on an interactive desktop session, describe it in the order notes for our engineers when you deploy the server. Our Ubuntu images include NVIDIA drivers, CUDA, Docker and the NVIDIA Container Toolkit.

On the 8-GPU server, rendering one frame range per GPU often scales better than splitting each frame across eight GPUs. With Blender, for example:

for gpu in 0 1 2 3 4 5 6 7; do
  CUDA_VISIBLE_DEVICES=$gpu blender -b shot.blend -o //frames/f_#### \
    -s $((gpu*30+1)) -e $((gpu*30+30)) -a -- --cycles-device OPTIX &
done
wait

Scene export, texture conversion and denoising passes also lean on the CPU and system RAM: the 8-GPU server pairs its GPUs with 128 cores and 1.5 TB of DDR5, while the single-GPU plans have 16 vCPUs with 188 GB of RAM on the RTX PRO 6000 and 128 GB on the L40S. Before a crunch, render a few representative frames on one GPU to measure peak memory and time per frame, then size the farm from those numbers rather than from guesses.

Sync assets in with rsync over SSH; inbound traffic is not counted. Keep caches and frames on local NVMe, about 15 TB raw on the 8-GPU server, 1.92 TB on a single RTX PRO 6000 and 1 TB on the L40S; see our NVMe storage guide. Outbound traffic includes 50 TB a month on the 8-GPU server and 20 TB on single-GPU plans. We do not back up local storage unless agreed in writing, so pull finished frames back every day.

Software licensing is your responsibility

We provide the hardware and the operating system, not render software licenses. Blender is free and open source. Commercial renderers and plugins such as Redshift and Octane need your own licenses, and you are responsible for following their terms, including any rules on rented, remote or cloud hardware and on the number of nodes. If you use floating licenses, the server must reach your license server: open only the ports it needs, as shown in our networking and firewall guide.

Booking render capacity for a crunch

The minimum term is one month, and you can cancel any month with no fee: the server runs until the end of the paid month. Plans renew automatically from your account balance, with an email reminder before each renewal, so book for the crunch and cancel from your dashboard before the next renewal date, as explained in renewals and cancellation. Servers are delivered in under 10 minutes after payment.

Rendering around the clock, the 8-GPU server at $10,490 a month works out to about $1.80 per GPU-hour; a single RTX PRO 6000 costs $1,390 a month and an L40S $890. For one weekend of rendering, an hourly render service will usually cost less; monthly capacity fits multi-week crunches and studios that render every day. Deploy the servers you need in one form: choose the configuration, create a free account or log in, and pay the first month from a prepaid balance topped up in crypto.