Nvidia L40S
Data centre GPU for AI and graphics workloads
The Nvidia L40S is a PCIe graphics processor built on the Ada Lovelace architecture. Launched in 2023, it targets generative AI training, inference, and high-end visualisation in enterprise servers.
- Uses 48 GB of GDDR6 memory for cost-effective capacity
- Consumes 350 W, exceeding previous PCIe power norms
- Supports FP8 precision for faster AI model training
- Lacks NVLink support for multi-GPU memory pooling
- Delivers up to 5X higher inference performance than A40
What it is
The Nvidia L40S is a graphics processing unit designed for enterprise data centres. It combines artificial intelligence compute capabilities with high-performance graphics rendering. Built on the Ada Lovelace architecture, the chip supports large language model training and inference alongside 3D visualisation tasks. Nvidia positions it as a universal accelerator for mixed workloads, replacing the earlier A40 series. The card fits into standard PCIe server slots and uses air cooling rather than liquid systems.
How it came about
Nvidia introduced the L40S at SIGGRAPH 2023 as an upgrade to the L40 and A40 GPUs. The company aimed to address growing demand for generative AI and industrial digitalisation tools. By increasing the power budget compared to its predecessor, the design allows for higher sustained clock speeds. This launch followed the broader adoption of Ada Lovelace silicon in consumer and professional markets, extending the architecture into heavy-duty server environments.
Inside the chip
The chip features fourth-generation Tensor Cores that support FP8 and FP16 data formats for faster AI processing. A Transformer Engine automatically adjusts precision between these formats to optimise memory use and speed. Third-generation RT Cores handle real-time ray tracing for graphics workflows. The unit includes 48 gigabytes of GDDR6 memory, which lowers cost compared to high-bandwidth memory solutions. It also contains an Optical Flow Accelerator to assist with frame generation and video upscaling via DLSS technology.
How it was made
Manufactured using the Ada Lovelace process, the L40S is built for durability in continuous operation. It meets Network Equipment-Building System Level 3 standards for telecom reliability. The physical card occupies two expansion slots and measures 4.4 inches in height by 10.5 inches in length. It connects via a PCIe Gen4 x16 interface, offering 64GB/s bidirectional bandwidth. Secure boot features provide a root of trust for data centre security requirements.
Where it went
System manufacturers such as Dell, Hewlett Packard Enterprise, and Lenovo integrate the GPU into OVX servers. These systems accelerate generative AI, large language model inference, and 3D rendering. Creative professionals use the hardware for virtual production and computer-aided engineering simulations. The card supports virtual GPU software, allowing multiple users to share its resources remotely. It also handles video encoding and decoding, including AV1 formats, for media workflows.
Versions and family
The L40S succeeds the Nvidia A40 and offers higher performance than the non-S L40 model. The primary difference from the L40 is a 350-watt power limit, up from 300 watts, which enables better throughput. Unlike some previous data centre chips, it does not support NVLink for pooling memory across multiple cards. This limitation keeps memory access confined to the single 48GB pool on each board. It sits below the Hopper architecture series in terms of interconnect capabilities.
Why it matters
The release marked a shift towards higher power consumption in PCIe form factors, breaking the previous 300W practical limit. By using GDDR6 memory instead of more expensive alternatives, Nvidia kept costs lower for visualisation tasks. The chip helps bridge the gap between pure AI accelerators and traditional graphics cards. Its support for OpenUSD workflows aids the development of metaverse and industrial digital twin applications.
Problems and reception
Observers noted the lack of NVLink support as a step back for users needing unified memory across multiple GPUs. Some critics expressed disappointment that mid-market data centre options remain limited in connectivity. However, the performance gains in tensor operations were recognised as significant, with some tests showing double the speed of the L40 in certain tasks. The use of generative fill in promotional materials drew attention to its AI capabilities.
Questions people ask about Nvidia L40S
What is the main difference between the L40S and the L40?
The L40S has a higher power limit of 350 watts compared to 300 watts on the L40. This extra headroom allows the GPU to maintain higher clock speeds, resulting in better performance for both AI and graphics tasks without changing the core architecture.
Does the L40S support NVLink for combining memory?
No, the L40S does not support NVLink. This means users cannot pool the memory of multiple cards into a single large address space. Each GPU operates with its own 48 GB of memory, which may limit complex scenes requiring more unified storage.
What type of memory does the L40S use?
The card uses 48 gigabytes of GDDR6 memory. This choice helps keep the cost of the card lower than models using High Bandwidth Memory (HBM), making it more attractive for remote visualisation and specific compute workloads.
Who manufactures servers with the L40S GPU?
Major system manufacturers including Dell, Hewlett Packard Enterprise, Lenovo, and Supermicro offer servers featuring the L40S. These OVX servers are designed to handle generative AI training, inference, and graphics-intensive enterprise applications.
Good reading and sources
There is no English Wikipedia article on Nvidia L40S yet, so the facts on this page come from the sources listed above: nvidia.com/en-us/data-center/l40s/, servethehome.com/nvidia-l40s-gpu-for-data-center-visualization-launche. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.