The NVIDIA L40S 48GB GPU is a next-generation accelerated computing solution designed for high-performance AI training, large language models (LLMs), graphics rendering, simulation, and virtual desktop infrastructure (VDI). Built on the advanced Ada Lovelace architecture, the L40S delivers up to 1.2X faster AI training and 2X faster inference performance than the previous NVIDIA L40. With 48GB GDDR6 memory, 18,176 CUDA cores, 568 Tensor cores, and 142 RT cores, it is engineered for the most demanding enterprise, data center, and cloud workloads.
Optimized for AI compute, visualization, Omniverse, Virtual Workstations, and Generative AI, the L40S accelerates end-to-end workflows—from model training to real-time rendering and deployment at scale. This GPU supports PCIe Gen 4.0, delivers exceptional energy efficiency per watt, and integrates seamlessly into enterprise servers and GPU-powered AI infrastructure.
| Specification |
Details |
| Model |
NVIDIA L40S 48GB |
| Architecture |
Ada Lovelace |
| CUDA Cores |
18,176 |
| Tensor Cores |
568 (4th Gen) |
| RT Cores |
142 (3rd Gen) |
| GPU Memory |
48GB GDDR6 ECC |
| Memory Bandwidth |
864 GB/s |
| Interface |
PCIe 4.0 x16 |
| Max Power (TDP) |
350W |
| NVENC / NVDEC |
4x NVENC + AV1 Encode / Decode |
| Multi-Instance GPU (MIG) Support |
Yes |
| Virtualization |
Supports NVIDIA vGPU, VDI & Omniverse |
| Form Factor |
Dual-slot GPU (Passive cooling for server use) |
| Cooling |
Passive – requires server airflow |
| Target Workloads |
AI, LLM Training & Inference, Rendering, VDI, Omniverse |
| Supported Frameworks |
TensorRT, CUDA, cuDNN, VMware, RedHat, Windows, Linux |
🧾 Compatible Part Numbers (Same Specs)
These part numbers represent the same NVIDIA L40S 48GB GPU with identical specifications.
-
900-2G133-0080-000
-
7WK28
-
900-2G133-0180-030
All part numbers listed above refer to the same NVIDIA L40S 48GB model with identical performance and specifications. Minor differences may relate to OEM packaging or system compatibility labeling.
❓ Frequently Asked Questions (FAQs)
Q1: What is the NVIDIA L40S mainly used for?
A: The L40S is ideal for AI training and inference (including LLMs), real-time rendering, Omniverse, visual computing, simulation, and VDI workloads.
Q2: How is the L40S different from the original NVIDIA L40?
A: The L40S delivers significantly higher performance, including nearly 2× AI inference throughput and improved training speeds, making it more suitable for Generative AI and LLM workloads.
Q3: Can the L40S be used for multi-user virtualization or VDI?
A: Yes - it supports NVIDIA vGPU, making it suitable for virtual workstations and enterprise VDI environments.
Q4: Does the L40S support AV1 encoding?
A: Yes - it includes next-gen AV1 encoding/decoding capabilities, ideal for streaming, media workloads, and virtual content creation.
Q5: Is external cooling required?
A: No external cooling is required; however, this GPU requires proper server airflow because it uses passive cooling.
