Mixing RTX 5090 and RTX 4090 for Rendering: VRAM Limits and Setup in Octane, Redshift, Cycles and Karma XPU
You can render with an RTX 5090 and an RTX 4090 together, with one firm limit:
- It works: Octane, Redshift, Cycles and Karma XPU all accept mixed NVIDIA GPUs in one machine.
- VRAM is per card: Every rendering GPU holds the full scene, so the pair renders in-core only up to 24 GB.
- Above 24 GB: Use the 5090 alone, or rely on out-of-core / system-memory fallback at a speed cost.
- Best use of both: Render one frame per GPU from the command line, as Maxon suggests for multi-GPU machines.
- On iRender: Identical GPUs in every server: 2x, 4x or 8x RTX 5090 (32 GB each) or RTX 4090 (24 GB each), never mixed.
Adding a new RTX 5090 next to an existing RTX 4090 is a common upgrade. It does speed up rendering, but it does not give you 56 GB of VRAM, and heavy scenes behave differently in each engine. This guide uses each vendor’s documentation to explain what to expect and how to get the most from a mixed pair.
Checked against Maxon, OTOY, Blender, SideFX and NVIDIA documentation, November 2026.
1. Can you mix an RTX 5090 and an RTX 4090?
Yes. In Puget Systems’ tests, mixed GPUs worked in both Octane and Redshift and the speed gain was in line with using matching cards; the catch they noted is that you are limited to the card with the least memory. The two cards differ mainly in memory size and bandwidth (NVIDIA RTX 5090, NVIDIA RTX 4090):
The two cards side by side
From NVIDIA's product pages.
| Spec | RTX 5090 | RTX 4090 |
|---|---|---|
| VRAM | 32 GB GDDR7 | 24 GB GDDR6X |
| Memory bandwidth | 1,792 GB/s | 1,008 GB/s |
| Architecture | Blackwell | Ada Lovelace |
| PCIe | PCIe 5.0 | PCIe 4.0 |
| Graphics card power | 575 W | 450 W |
2. Why don't 32 GB and 24 GB add up?
GPU renderers copy the whole scene to every GPU that renders it, so a mixed pair renders in-core only up to the smaller card’s 24 GB.
A 26 GB scene on a mixed pair
Each GPU receives its own full copy of the scene. Illustrative example; real limits depend on the engine and on VRAM already used by the system.
│ copied to every rendering GPU
├──▸ RTX 5090 · 32 GB → 26 GB scene fits, about 6 GB left
└──▸ RTX 4090 · 24 GB → 26 GB does not fit
├─ out-of-core on → data paged from system RAM, slower
└─ out-of-core off → the card cannot render this scene
A note on sources about VRAM: Maxon's support article states that Redshift does not combine the VRAM when using multiple GPUs. Maxon's 3D rendering product page also lists "multi-GPU memory pooling" and says Redshift shares out-of-core memory across multiple cards. This article follows the support article: plan for every GPU holding the scene in its own VRAM. Check Maxon's latest documentation for your Redshift version.
3. How do Octane, Redshift, Cycles and Karma XPU handle a mixed pair?
What each vendor documents
Quoted or summarized from each vendor's own documentation.
| Engine | VRAM across GPUs | When a GPU runs out |
|---|---|---|
| Redshift | Not combined (an 8 GB + 12 GB pair is not 20 GB) | Out-of-core rendering, so the largest scene is not limited by the weakest GPU |
| Octane | Limited to the VRAM of the smallest GPU | Out-of-core keeps textures and geometry in system RAM; OTOY recommends disabling GPUs without enough VRAM |
| Cycles | Each GPU can only access its own memory | CUDA, OptiX, HIP and Metal try system memory automatically, with a speed cost |
| Karma XPU | Each device addressed separately; memory not pooled | If a device fails, other devices (including the CPU) finish the frame |
4. How to use both cards well (step by step)
Step 1: Render heavy scenes on the RTX 5090 only
When a scene is over 24 GB, untick the RTX 4090 for rendering so the scene stays in-core on the 32 GB card:
Redshift (C4D) .. Edit › Preferences › Renderer › Redshift › Compute Devices
untick the RTX 4090
Octane ......... Preferences › Devices › Render: RTX 5090 only
(the RTX 4090 can stay ticked for Denoise)
Cycles ......... Preferences › System › Cycles Render Devices › OptiX
untick the RTX 4090
Karma XPU ...... KARMA_XPU_DEVICES = list of device IDs to enable
Menu sources: Maxon, OTOY, Blender, SideFX. Karma XPU shows device IDs in its render stats.
Step 2: Or render one frame per GPU
Maxon notes that giving each job a subset of GPUs lets you render several frames at once on one machine for better scaling (Maxon). Each card then works at its own speed and keeps its own memory limit. Run the two commands in two terminals:
Sources: Maxon command-line GPU selection, Redshift environment variables, OTOY command line, Blender command line. Check which index belongs to which card in each app before running.
Step 3: Keep out-of-core on for large scenes
Quick check: during a test frame, nvidia-smi shows each job on its own GPU and memory below that card's total.
5. Mixed pair at home or identical GPUs on a farm: SaaS or IaaS?
A SaaS render farm takes the scene as a job and manages the hardware for you, which suits standard scenes and supported plugins. An IaaS render farm such as iRender rents you a whole server with identical GPUs, 2x, 4x or 8x RTX 5090 (32 GB each) or RTX 4090 (24 GB each), so every card has the same memory limit and you choose the devices yourself. See SaaS vs IaaS render farm and multi-GPU render engines for Cinema 4D.
Multi-GPU rendering on iRender servers with identical GPUs.
On iRender you rent a whole server with 2x, 4x or 8x RTX 5090 or RTX 4090 GPUs, with Cinema 4D and Redshift pre-installed; Prime members get activated Cinema 4D and Redshift licenses. Real render times are in our Redshift benchmark, and plans are on the Cinema 4D Redshift render farm page.
6. Mixed GPU rendering FAQ
Can I use an RTX 5090 and an RTX 4090 together for rendering?
Yes. Octane, Redshift, Cycles and Karma XPU all work with mixed NVIDIA GPUs. The pair is faster than either card, but every card must hold the scene.
Do an RTX 5090 and RTX 4090 combine to 56 GB of VRAM?
No. Each rendering GPU keeps its own copy of the scene, so the in-core limit is the smaller card: 24 GB. Maxon, OTOY, Blender and SideFX all document that VRAM is not pooled.
What happens when my scene is bigger than 24 GB?
Redshift and Octane can use out-of-core memory and Cycles falls back to system memory, all slower than in-core. Karma XPU lets the other devices finish the frame if one GPU fails. Rendering on the RTX 5090 alone keeps a scene up to about 32 GB in-core.
Does the slower RTX 4090 slow down the RTX 5090?
The pair is still faster than one card; Puget Systems found mixed cards scaled in line with expectations in Octane and Redshift. Test your own scene, because results vary by engine and scene.
How do I render one frame per GPU?
Start two command-line renders, each limited to one GPU: -redshift-gpu for Cinema 4D with Redshift, -g for Octane Standalone, a device setting via –python-expr for Blender, or KARMA_XPU_DEVICES for Karma XPU, each with its own frame range.
Does iRender mix GPU models in a server?
No. Each iRender server has identical GPUs: 1, 2, 4 or 8 RTX 4090 cards, or 1, 2, 4 or 8 RTX 5090 cards.
