GPU render server with RTX PRO 6000: scene memory, multi-GPU scaling and renderer support

What decides a GPU render server

A GPU render server, the render node of a render farm, is sized by two numbers: how many cards it holds and how much memory each card has. GPU renderers such as Blender Cycles, V-Ray GPU, Redshift and Octane distribute the samples of one frame across all cards in the node, so render time falls almost in proportion to the card count. Memory behaves differently, because each card needs its own copy of the scene, so the largest scene a node can render at full speed is set by the memory of one card, 96 GB on the RTX PRO 6000 Blackwell, and not by the sum of all cards.

The RTX PRO 6000 Server Edition carries 188 fourth-generation RT Cores on the same GPU as the two desktop editions, and the renderers use them through NVIDIA OptiX or their own RTX modes. NVIDIA rates the Workstation Edition at 380 TFLOPS of RT Core performance and 125 TFLOPS of FP32, and the Server Edition at 355 and 120 TFLOPS. The rest of the server follows from the card count: CPU cores, system memory, power and cooling. Licences per render machine come on top, and each renderer vendor sets its own terms.

Renderer support for GPUs, several cards and large scenes

Each renderer documents its own GPU path, its behaviour with several cards and what happens when a scene does not fit. The table below summarises what the vendors’ documents state as of October 2026.

Blender 5.2 LTS manual (GPU rendering); Chaos support articles of 26 September 2025, 30 October 2025, 26 May 2026 and 15 July 2026; Maxon’s Redshift page and its documentation on out-of-VRAM situations; OTOY’s out-of-core documentation and hardware guide (updated 28 October 2025); Epic’s Unreal Engine 5.8 path tracer documentation.

The fallback modes keep a render running, at a cost. The Blender manual says Cycles moves to system memory when GPU memory is full, which “has a performance impact, but will usually still result in a faster render than using CPU rendering”. OTOY says a speed trade-off is expected with out-of-core data, and that RTX acceleration in Octane “can not be used unless your whole scene fits into the VRAM on the card”. Chaos states that the entire scene, with geometry, textures and buffers, must fit into GPU memory in V-Ray’s CUDA and RTX modes. Its CUDA x86 mode uses system RAM instead, on the CPU.

Scene memory per card: why 96 GB is the budget

Several cards do not form one large memory pool. The Blender manual answers the question of whether multiple GPUs increase available memory with “Typically, no, each GPU can only access its own memory”, and names NVIDIA GPUs connected with NVLink as the exception. Chaos advises that, with cards of different memory sizes, the scene should fit on the GPU with the least memory. OTOY’s hardware guide, updated 28 October 2025 with passages written in November 2020, says that “in most cases Octane does not pool (or combine) VRAM from multiple cards”. For the RTX PRO 6000, NVIDIA’s product brief for the Server Edition (June 2025) lists NVLink as “Not supported”. Chaos adds that NVLink support is found on larger GPUs before the Ada generation.

A node with four RTX PRO 6000 cards therefore renders the same maximum scene size as one card, at up to four times the speed. The RTX 6000 Ada and the L40S each have 48 GB, so scenes between 48 and 96 GB stay in GPU memory only on the Blackwell card and keep the faster path. Denoising adds to the memory load, and the Blender manual says denoising on the GPU “requires additional GPU memory” and that the option can be disabled when large scenes need more GPU memory.

The input to size from is the peak GPU memory of the heaviest shot, read from the renderer’s statistics or nvidia-smi on the current machine. Add the denoiser and a margin for scene growth, then compare the total with 96 GB.

We build VDI and rendering nodes with RTX PRO Blackwell, L40S or L4 cards, sized by seats or frames per card. Tell us your renderer and the peak memory of your heaviest scene, and we size the node around it.

Multi-GPU scaling and cards per render node

For V-Ray GPU, Chaos published scaling tests in a support article updated on 26 September 2025, run on eight RTX 3090 Ti cards. Across its CUDA and RTX tests it recorded an average of 97% with 4 GPUs and 91% with 8 GPUs. Chaos also says V-Ray GPU officially supports a maximum of 4 GPUs, although all GPUs the system recognises are usable. The Redshift pages we could read give no scaling figure, and Maxon’s plan covers “up to 8 GPUs+ on a Single Computer”. OTOY adds the OctaneBench scores of all cards to a system score and describes the result as “More cards = more rendering speed.”

The Unreal Engine path tracer is the exception. Epic’s Unreal Engine 5.8 documentation supports multiple GPUs through NVIDIA’s SLI technology, with the cards connected by NVLink bridges, SLI enabled in the NVIDIA Control Panel and the editor started with -MaxGPUCount=N. The RTX PRO 6000 has no NVLink, so by Epic’s requirement one path tracer process on such a node renders on one card. In our reading, a node with several cards then serves Unreal through one render job per card, set up in the render manager.

The CPU and system memory scale with the cards. Chaos recommends at least 6 physical cores per GPU, so a node with four cards needs 24 cores and one with eight needs 48. It also recommends system memory equal to or greater than twice the GPU memory. For four RTX PRO 6000 cards with 384 GB of GPU memory, that is 768 GB of system RAM. OTOY’s hardware guide asks for three to four times more RAM than GPU memory, from passages written in 2020. Out-of-core modes in Octane, Redshift and Cycles also hold scene data in system RAM, which is one more reason not to size the RAM down.

Published render benchmarks for RTX PRO cards

Blender Open Data collects Cycles results submitted by users. Its score is, in the site’s words, “the estimated number of samples per minute, summed for all benchmark scenes”, measured on one CPU or GPU device. The table shows the median scores for Blender 4.5.0 with OptiX.

Blender Open Data, query for Blender 4.5.0, compute type OptiX, grouped by device, read on 10 October 2026; scores rounded to whole numbers. User submissions, single device per result.

By these medians, the Server Edition scores about 7% below the Workstation Edition and the Max-Q about 12% below, our arithmetic. The Workstation Edition scores 57% above the RTX 6000 Ada, and the Server Edition about twice the L40S. Rows with fewer than ten results give only an order of magnitude. Each result covers one card, so a node’s throughput follows from the vendors’ scaling figures above, not from this table. Our RTX PRO 6000 Blackwell vs RTX 6000 Ada comparison lists the published V-Ray and Blender pairs for both generations.

RTX PRO 6000 for rendering: Workstation, Max-Q or Server Edition

All three editions have the same GPU and 96 GB. The Workstation Edition runs at 600 W with flow-through cooling, and makers list one or two per tower. The Max-Q runs at 300 W with a blower, and NVIDIA states up to four per system. Our guide to four RTX PRO 6000 Max-Q cards in one workstation covers the towers, power supplies and heat for that build. It suits a studio without a server room that wants a shared render machine next to the artists.

The Server Edition is passively cooled and belongs in a rack server qualified for it. NVIDIA’s product brief gives 600 W as the default maximum, or 450 W with a power cable strapped for that mode, and Lenovo caps it at 450 W to fit four cards in its SR650a V4. A capped card renders more slowly, so state the power level when you compare quotes. The Server Edition is also the only edition with vGPU. Our comparison of the Workstation, Max-Q and Server Edition has the full specifications and the power cable details.

Virtual workstations for artists with vGPU

Where artists work remotely or the studio wants the GPUs in the server room, the Server Edition can host virtual workstations as well as render. Lenovo’s product guide lists vPC, vApps and RTX Virtual Workstation (vWS) as supported vGPU software. As our guide summarises NVIDIA’s documents, vWS is licensed per concurrently active user, and NVIDIA’s sizing guide places two to four heavy 3D users on one card. Time-sliced vGPUs share the card’s compute, so a final render inside one VM slows its neighbours. Our guide to MIG and vGPU on the RTX PRO 6000 has the profiles, seat counts and licence rules.

Render node licences

Licensing per render machine differs by vendor, so check it before the card count is fixed. Chaos tells V-Ray users to add one render node licence for each additional render machine. OTOY states that all Octane Studio+ subscriptions include 10 or more render node licences, “with no GPU limits per machine”. Maxon includes Redshift with every Cinema 4D subscription and lists “Multiple Instances up to 8 GPUs+ on a Single Computer” in its plan. Terms change between product versions, so read those of your own subscription.

Render server configurations by studio size

The configurations below are our own estimates as starting points for sizing, not measured results or vendor recommendations. The final build follows from the renderer, the heaviest scene and the frame volume.

Our estimates, not vendor figures, using Chaos’s 6 cores per GPU and twice the GPU memory in RAM; seat counts per card from NVIDIA’s vGPU sizing guide as summarised in our MIG and vGPU guide.

NVIDIA’s RTX PRO Server reference design holds eight Server Edition cards. For V-Ray, two nodes with four cards stay within the configuration Chaos supports officially. Four cards at 600 W draw 2.4 kW before the processors and fans, so the rack position needs a matching feed.

We check the rack, power and airflow before we quote. Describe your studio, renderer and frame volume in the form below, and we reply with a render node configuration within one business day.

What we supply

We build AI servers and rendering nodes to order with RTX PRO Blackwell, L40S or L4 cards, assembled and burn-in tested, with manufacturer warranty on every component and delivery anywhere in the EU. We supply the RTX PRO 6000 in all three editions, the RTX PRO 5000 and 4500 and the other professional NVIDIA GPUs on one EU contract and invoice, with vGPU licences on the same invoice as the hardware. We also build on your own chassis or with parts you already own, after a compatibility check of the platform, power and cooling. Operating system, drivers and CUDA are installed on request.

FAQ