NVIDIA L40S 48GB – Ada Lovelace Data Center GPU

Description

The NVIDIA L40S is a versatile enterprise GPU designed to consolidate multiple accelerated workloads onto a single data-center platform. Based on the Ada Lovelace architecture, it provides a combination of GPU compute, AI acceleration, ray tracing, visualization, and professional media-processing capabilities.

With 48GB of GDDR6 memory with ECC, the L40S is well suited to workloads that require substantial GPU memory capacity, including generative AI models, LLM inference, rendering scenes, engineering visualization, simulation, and large professional datasets.

The GPU contains 18,176 CUDA cores, 568 fourth-generation Tensor Cores, and 142 third-generation RT Cores. NVIDIA rates the L40S at up to 91.6 TFLOPS FP32, 212 TFLOPS RT Core performance, and 1,466 TFLOPS FP8 Tensor performance with sparsity.

The fourth-generation Tensor Cores and NVIDIA Transformer Engine provide hardware acceleration for modern AI workloads, including FP8 and FP16 operations. This makes the L40S particularly suitable for generative AI, LLM inference, AI model development, and data-science workloads.

For graphics and visualization, the third-generation RT Cores accelerate real-time ray tracing and rendering. The L40S can therefore support professional visualization workloads alongside AI and compute applications, making it a versatile choice for enterprise data centers.

The GPU also incorporates three NVENC encoders and three NVDEC decoders, including AV1 encode and decode, supporting professional video production, streaming, transcoding, and media-processing workloads.

3. Complete Technical Specification

Specification	Details
Manufacturer	NVIDIA
Product	NVIDIA L40S
Product Class	Data Center GPU
Architecture	NVIDIA Ada Lovelace
CUDA Cores	18,176
Tensor Cores	568, 4th Generation
RT Cores	142, 3rd Generation
GPU Memory	48GB
Memory Type	GDDR6
Memory ECC	Yes
Memory Bandwidth	864GB/s
System Interface	PCIe Gen4 x16
PCIe Bandwidth	64GB/s bidirectional
FP32 Performance	91.6 TFLOPS
TF32 Tensor Performance	366 TFLOPS with sparsity
FP16 Performance	733 TFLOPS with sparsity
BF16 Performance	733 TFLOPS with sparsity
FP8 Performance	1,466 TFLOPS with sparsity
RT Core Performance	212 TFLOPS
INT8 Tensor Performance	Up to 1,466 TOPS with sparsity
INT4 Tensor Performance	Up to 1,466 TOPS with sparsity
Form Factor	4.4″ H × 10.5″ L
Slot Width	Dual-slot
Cooling	Passive
Maximum Power Consumption	350W
Power Connector	16-pin
Display Outputs	4 × DisplayPort 1.4a
Video Encoders	3 × NVENC
Video Decoders	3 × NVDEC
AV1	Encode and Decode
vGPU Support	Yes
MIG Support	No
NVLink Support	No
Secure Boot	Yes, with Root of Trust
NEBS	Level 3 Ready
Primary Market	Data Center / Enterprise
Typical Workloads	AI, LLM, rendering, visualization, video

The above specifications are based on NVIDIA’s current official L40S documentation.

4. Applications

The NVIDIA L40S is designed for a broad range of accelerated data-center applications:

Generative AI
Large Language Model (LLM) inference
LLM training
AI model development
AI inference services
Machine learning
Deep learning
Computer vision
Natural language processing
3D rendering
Real-time ray tracing
Professional visualization
Engineering visualization
Digital twins
CAD/CAE workloads
Scientific visualization
Video production
Video transcoding
Streaming
AV1 media processing
Virtual workstations
Cloud GPU services
Enterprise data-center acceleration

NVIDIA positions the L40S specifically as a multi-workload accelerator spanning generative AI, LLM workloads, graphics, rendering, and video.

5. Compatibility

Hardware Compatibility

The L40S is designed for qualified data-center servers featuring:

PCIe Gen4 x16
Dual-slot GPU support
Adequate 350W-class power delivery
Compatible 16-pin GPU power connection
Sufficient chassis airflow and thermal design
Passive GPU cooling with server-level airflow

Because the L40S uses passive cooling, it relies on the host server’s airflow system rather than an onboard GPU fan.

Software Compatibility

The L40S supports NVIDIA’s professional and enterprise software ecosystem, including:

NVIDIA CUDA
NVIDIA AI Enterprise
NVIDIA vGPU software
NVIDIA RTX technologies
NVIDIA OptiX
TensorRT
NVIDIA Omniverse
Professional visualization applications
AI and machine-learning frameworks

NVIDIA confirms vGPU software support for the L40S.

Important Compatibility Notes

The L40S is a data-center passive GPU, not a conventional consumer gaming graphics card. It should be installed in a server or workstation platform specifically designed to provide sufficient airflow and power.

It also does not support NVLink or MIG, so these capabilities should not be listed as features when comparing it with GPUs such as the H100 or B200.