An eight-GPU NVIDIA RTX PRO 6000 Blackwell server holds 768 GB of GPU memory. For most GPU renderers, the number that decides whether a scene fits is 96 GB, the memory on one card. This post works through what that means for rendering, VFX and transcoding, using figures from NVIDIA and from renderer documentation.

In October our team ran the same arithmetic for language models in 8 x 96 GB: what fits on an RTX PRO 6000 Blackwell node. Media workloads use the memory differently, so the answers change.

Why a GPU renderer sees 96 GB and not 768 GB

The RTX PRO 6000 Blackwell Max-Q has 96 GB of GDDR7 with ECC, 1,792 GB/s of memory bandwidth and a 300 W board power limit. It connects to the host over PCIe 5.0 x16.

Most GPU renderers split the frame across cards, and each card renders from its own copy of the scene. The Blender manual puts it plainly: adding GPUs typically does not increase available memory, because “each GPU can only access its own memory.” Eight cards render faster than one. They do not hold a scene eight times larger.

So the sizing question for a render node is per card: does the scene, with its textures, geometry, acceleration structures and frame buffers, fit in 96 GB?

256 MB per 8K texture: the arithmetic for one scene

The Blender manual gives a useful unit: one 8K texture takes 256 MB. That is 8,192 × 8,192 pixels at four 8-bit channels, uncompressed. A 16-bit texture of the same size takes twice as much, and compressed or tiled formats take less.

As an illustration, a scene with 300 uncompressed 8-bit 8K textures needs about 75 GB for textures alone, before geometry, displacement, volumes and render buffers. On a 96 GB card it fits with room left for geometry. On a 48 GB card it does not fit, and the renderer has to move data to system memory or fail.

The real figure for any project comes from the renderer’s own memory report. Redshift, for example, reports geometry memory and the texture cache separately, which shows which of the two is growing.

What out-of-core buys, and what it costs

Most production GPU renderers can keep part of the scene in system memory when it does not fit on the card:

  • Redshift supports out-of-core geometry and textures. Its documentation says textures do not need to fit entirely in VRAM: it uploads only the tiles it needs and uses lower-resolution MIP maps for distant textures.
  • Octane can keep textures and geometry in system RAM. Its documentation says “some tradeoff in the rendering speed is expected,” and CPU load rises because the CPU serves the memory requests.
  • Cycles falls back to system memory automatically on CUDA and OptiX devices, with a performance cost.

Out-of-core makes a large scene renderable. It does not make it render at on-card speed, because data that crosses PCIe is slower than data already in GDDR7. A card that holds the whole working set avoids the trade-off. On a dedicated server, the system memory used for out-of-core belongs to your jobs only.

Where 48 GB stops being enough

The NVIDIA L40S and RTX A6000 both have 48 GB per GPU. For many scenes that is enough. It runs short in three common cases:

  • texture-heavy assets with many 8K or 16-bit maps, as in the arithmetic above
  • heavy displacement, volumes or simulation caches loaded for a single frame
  • higher output resolutions with many AOVs, where frame buffers grow with resolution

The RTX PRO 6000 doubles the per-card budget to 96 GB. That moves the threshold at which a scene goes out-of-core, which matters more for render time than the difference in core count.

32 NVENC and 32 NVDEC engines per node

Transcoding uses a different part of the GPU. Each RTX PRO 6000 Blackwell has four ninth-generation NVENC encoders and four sixth-generation NVDEC decoders, which gives an eight-GPU node 32 of each. NVIDIA’s support matrix lists concurrent encode sessions on the card as unrestricted.

The codec support is the larger change for media pipelines. NVIDIA lists H.264 and HEVC 4:2:2 encoding and AV1 encoding for the RTX PRO 6000 Blackwell. The L40S has three eighth-generation NVENC encoders with AV1 but no HEVC 4:2:2. The RTX A6000 has one seventh-generation NVENC with neither 4:2:2 nor AV1. For teams delivering 10-bit 4:2:2 masters or AV1 streams, that decides whether encoding runs on the GPU at all.

Four 24 GB instances per GPU

The RTX PRO 6000 Blackwell supports Multi-Instance GPU. One card can run as four 24 GB instances, two 48 GB instances or one 96 GB GPU. On an eight-GPU node, that is up to 32 isolated instances for work that does not need a full card, such as encoding queues, preview renders or lookdev sessions, alongside full-card instances for final frames.

A sizing checklist for a render node

  • Peak memory per frame from your renderer’s log, not the scene file size
  • Whether that peak fits in 96 GB per card with headroom for buffers
  • How often jobs go out-of-core today, and what it adds to render time
  • Codecs and chroma formats you deliver: 4:2:2, 10-bit, AV1
  • Whether part of the queue could run on 24 GB or 48 GB MIG instances

Where these servers run

RTX PRO 6000 Blackwell servers are in service in Montreal, with allocation open, as whole servers on 36 to 60 month terms. Every server is built to our delivery specification and passes burn-in and acceptance before handover. The media and entertainment page lists the servers we size for rendering, VFX and transcoding.

Request a quote with the hardware, number of servers, term and start date, or talk to an engineer.