GPU servers with NVIDIA L40S

The NVIDIA L40S is a data-centre card built on the Ada Lovelace architecture: 48 GB of GDDR6 memory with ECC, tensor cores with FP8 support, RT cores for ray tracing and hardware codecs with AV1. It sits between low-power inference cards and training accelerators: twice the memory of 24 GB cards, at a noticeably lower price than an H100. The table above lists cloud plans and dedicated servers with this card, from a vGPU slice with a few gigabytes of memory to four-card builds. If you are looking at an L40S server rental, first check what you actually get: the whole card or a part of it.

What fits in 48 GB of video memory

Memory is the main reason to pick the L40S. A rule of thumb for language models: about 2 GB per billion parameters in FP16, about 1 GB in 8-bit quantization and 0.5–0.6 GB in 4-bit. On top of that you need headroom for the KV cache: the longer the context and the more parallel requests, the more memory it takes.

• 7–13B models in FP16 fit with room to spare: a 13B model takes about 26 GB, leaving space for long context and batched requests.
• 30–34B models in 8-bit need about 34–36 GB for weights. That is a workable setup for one card with moderate context.
• 70B models fit only in 4-bit quantization: the weights take about 40 GB, leaving a few gigabytes for context. Enough for a single user and short conversations; for a service under load, two cards are better.
• Fine-tuning. LoRA on 7–13B models and QLoRA up to roughly 30B fit on one card. Full training of even a 7B model with the Adam optimizer needs more than 48 GB.

If your model is under 8B parameters and needs up to 24 GB, compare prices for L4 servers: they are more economical for that load. For training large models from scratch, the H100 and A100 are a better fit: their HBM memory has far higher bandwidth and the cards are linked through NVLink.

Images, rendering and video

This is where the L40S is stronger than in training. Stable Diffusion and Flux generate images on it in seconds, and 48 GB lets you work at high resolution, keep several models in memory at once and train your own LoRAs without offloading weights. RT cores speed up rendering in Blender Cycles via OptiX and in other GPU renderers, and the large memory holds heavy scenes with high-resolution textures. Hardware NVENC and NVDEC with AV1 support handle video transcoding without loading the CPU.

vGPU slice, full card and ×2, ×4 builds

The cheapest rows in the table are not full cards. Plans named "L40S-4Q", "-8Q", "-12Q", "-16Q" give you a share of one card through NVIDIA vGPU: the virtual machine gets 4, 8, 12 or 16 GB of video memory and a portion of compute time. The letter Q stands for the virtual workstation profile.

• A 4–8 GB slice suits a remote workstation: CAD, 3D modelling, hardware-accelerated editing. You will not run a language model of useful size in 4 GB.
• A 12–16 GB slice handles image generation and small 4-bit models, but compute units are shared with neighbours and speed is not guaranteed.
• A full 48 GB card is what you need for inference of large models, fine-tuning and steady performance.
• ×2 and ×4 builds give 96 and 192 GB in total. The L40S has no NVLink, so cards talk over PCIe 4.0, and splitting one model across cards is slower than on NVLink accelerators. Several independent model copies, one per card, scale almost linearly though.

The L40S does not support MIG: the card is shared only through vGPU profiles.

Driver, CUDA and what to check in the image

• Compute capability 8.9. You need CUDA 11.8 or newer; for FP8 and current PyTorch releases pick an image with CUDA 12.
• On vGPU slices the driver is installed by the host: the machine runs a vGPU guest driver, and you should not install a regular driver over it. Ask for its version in advance.
• nvidia-smi output after first boot shows the model, the allocated memory and the CUDA version; compare them with the plan.

Hourly or monthly billing and price range

In the catalogue a GPU server with the L40S costs roughly $115 a month for a vGPU slice up to about $4,500 for the most powerful multi-card configurations; the median is around $960. Fewer than half of the plans bill by the hour. Hourly billing pays off for one-off jobs: a fine-tuning run, rendering a project, testing a model before launch. If the card is busy most of the month, a monthly plan or a dedicated GPU server is usually cheaper. How to compare such plans is covered on the hourly billing page. Check the data centre location too: for personal data of EU residents a data centre in the EU keeps GDPR compliance simple, and a nearby location cuts latency for interactive use and speeds up dataset downloads. Payment methods per plan are listed in the table; most international hosts take regular bank cards.