An A40 GPU cloud instance is a quiet favourite for two kinds of work: headless 3D rendering and mid-size AI inference that needs more than 24 GB of memory. The A40 is a datacenter card with 48 GB-class memory, built for servers rather than desktops. On Cloudj you find one by opening the deploy page with the A40 selected and narrowing the list of independent host machines by country, for example US, DE, NL. This article explains where the A40 tends to sit next to the L40S and the RTX A6000, how to use the Datacenter only switch, and how to run the two jobs it is most often rented for.
For how location works across all regions, the GPU cloud by location guide is the starting point.
Where the A40 sits among 48 GB cards
Three cards often come up together when 24 GB is not enough. They share the 48 GB class but differ in generation and intent.
| A40 | L40S | RTX A6000 | |
|---|---|---|---|
| Class | Datacenter | Datacenter | Workstation |
| Generation | Ampere | Ada Lovelace | Ampere |
| Memory | 48 GB class | 48 GB class | 48 GB |
| Typical use | Rendering, visual computing, mid-size inference | Newer-generation inference and training, rendering | Rendering, inference, development |
In broad terms, the A40 and the RTX A6000 are siblings of the same generation, one designed for servers and one for workstations. The L40S is a newer generation and is usually the stronger performer, which often shows in its price. Check the specifications on each model page, and compare live "from" prices on GPU cloud rather than relying on general rules. For a deeper side-by-side, see L40S vs A100 vs RTX 6000 Ada.
All three need a higher account level than a new account has. If you try to deploy one too early, the console says the GPU class is not available at your account level, and you can contact Cloudj support from there.
Searching by region
Open the deploy page with the A40 and enter the codes that suit your users and data:
USfor North American work or for batch jobs where location barely matters.DE, NLfor European teams, with more EU codes if the list is short.US, DE, NLfor a wide first look.
Each row reports its country at country level only. Stock changes during the day, so if nothing is listed, try the L40S or RTX A6000 with the same codes, widen the codes, or check later.
The Datacenter only switch
The A40 is a datacenter card by design, but the Datacenter badge in the results describes where the host machine sits, not what kind of card it has. If you want hosts located in a data centre, turn on the Datacenter only switch. The list gets shorter and prices are often higher, but you filter out machines outside data centres.
Even with the switch on, each machine is an independent host with no hardware SLA. Keep backups either way. The datacenter vs independent hosts guide explains the trade-off in more depth.
Job 1: headless rendering with Blender
There is no graphical desktop on GPU instances, so rendering is done from the command line. Upload your .blend file and assets, then render with Cycles on the GPU:
blender -b scene.blend -E CYCLES -o //renders/frame_#### -s 1 -e 240 -a -- --cycles-device OPTIX-b runs Blender without a UI, -s and -e set the frame range, and -a renders the animation. Check that the GPU is in use with nvidia-smi in a second SSH session. The Blender command line documentation lists every argument.
Install Blender over SSH or bring it in a custom Docker image. For European studios, the Blender rendering on GPUs in Europe guide covers file transfer and frame batching.
Job 2: mid-size inference
With 48 GB, an A40 can hold models that do not fit on a 24 GB consumer card, or the same model with a longer context and more concurrent requests. Typical cases:
- a mid-size open model at 8-bit, or a larger one at 4-bit;
- a multimodal model plus an embedding model on the same card;
- image generation pipelines that keep several large models loaded.
Install your serving stack yourself, for example with pip over SSH, and bind it to localhost. Reach it through an SSH tunnel or put a VPS in front for a public API. Measure memory with a realistic load before you commit to a long run.
When the A40 is not the right choice
- If your model fits in 24 GB, a consumer card is usually simpler and does not need a higher account level.
- If you need the newest generation features or the most throughput, look at the L40S or larger datacenter cards.
- If you need a guaranteed machine with an SLA, a live inventory of independent hosts is not the right tool.
FAQ
Is the A40 always in a data centre?
The card is designed for servers, but the Datacenter badge refers to the host. Turn on Datacenter only if that matters to you.
Does the A40 need a higher account level?
Yes. Datacenter and workstation GPUs need a higher level; the console tells you and you can contact support from there.
Can I use the A40 for video encoding?
It has hardware video engines, and ffmpeg with NVENC runs headless. Check codec support on NVIDIA's documentation for your format.
Where are A40 prices?
Live on the A40 page and GPU cloud.
Try it
Open the deploy page with the A40, enter US, DE, NL, and decide whether you want the Datacenter only switch on. Compare it with the L40S and RTX A6000 for your job, and keep backups of your scenes, weights and outputs.