Server with NVIDIA A40: 48 GB for inference, rendering and virtual workstations
The NVIDIA A40 is an Ampere data-centre card. According to the NVIDIA specification, it has 48 GB of GDDR6 with error correction (ECC) and 696 GB/s of bandwidth, 10,752 CUDA cores, passive cooling and a power draw of up to 300 W. Two cards can be joined with NVLink (112.5 GB/s), giving 96 GB as a pair. The card supports NVIDIA vGPU, so hosts offer it both as a whole card and in slices.
What it is good for
- Language models that do not fit in 24 GB: 30–34B parameters in 8-bit, 70B in 4-bit quantization with a short context.
- Rendering and 3D: second-generation RT Cores and plenty of memory for heavy scenes.
- Virtual workstations for CAD and video editing via vGPU profiles.
The A40's memory is slower than on gaming cards: when generating language model output, the RTX 4090 with its 1 TB/s is faster, as long as the model fits in 24 GB. A newer and faster card in the same 48 GB class is the L40S.
Whole card or vGPU slice
The cheapest rows in the table are slices of one card with 2, 4 or 8 GB of video memory. They suit a remote workstation and light graphics, but not language models. Inference and rendering need a whole 48 GB card, so check the memory size in the plan. The A40 has no MIG: it can only be shared through vGPU.
Monthly and hourly cost
In the catalogue as of 17 November 2026, the median price of a server with one whole A40 is about $1,050 a month, with half of the plans between $600 and $1,100. Over 730 hours that comes to about $1.44 an hour. There are both cloud plans and dedicated servers; only a few hosts offer hourly billing. If you need the card for a few hours, compare hourly plans and read how to work out the cost of an hour. How much memory a model takes is covered in the article on running your own language model on a server.