NVIDIA L40 48GB – Ada Lovelace Data Center GPU
The NVIDIA L40 48GB is a high-performance data-center accelerator built on the NVIDIA Ada Lovelace architecture. It combines 18,176 CUDA cores, 568 fourth-generation Tensor Cores, 142 third-generation RT Cores, and 48GB of ECC-protected GDDR6 memory.
With 864GB/s memory bandwidth, PCIe 4.0 x16 connectivity, and a 300W maximum board power, the L40 is designed for enterprise AI, inference, professional visualization, 3D rendering, virtual workstations, simulation, and GPU-accelerated computing.
Description
The NVIDIA L40 is a professional data-center GPU engineered to accelerate a wide range of enterprise workloads, including artificial intelligence, deep learning inference, graphics-intensive applications, 3D rendering, scientific visualization, and virtualized workstation environments.
Based on NVIDIA’s Ada Lovelace architecture, the L40 incorporates 18,176 CUDA cores, 568 fourth-generation Tensor Cores, and 142 third-generation Ray Tracing Cores. This combination provides substantial parallel compute capability while enabling hardware acceleration for AI and real-time ray tracing.
The L40 includes 48GB of ECC-protected GDDR6 memory, providing the capacity required for large AI models, complex 3D scenes, high-resolution visualization datasets, engineering applications, and professional graphics workloads.
Its 384-bit memory interface delivers up to 864GB/s of memory bandwidth, allowing large datasets and graphics assets to move rapidly between the GPU’s memory and processing resources.
For AI workloads, the L40 supports Tensor Core acceleration and modern reduced-precision computing, including FP8, BF16, FP16, TF32, INT8, and FP32 workloads. This makes it suitable for AI inference, deep-learning applications, recommendation systems, computer vision, and other machine-learning workloads.
The L40 also includes dedicated video-processing hardware with two NVENC encoders and two NVDEC decoders, including AV1 encoding and decoding capabilities. This provides hardware acceleration for professional media processing, video streaming, content creation, and virtual production.
Unlike the workstation-oriented RTX 6000 Ada, the L40 uses a passive cooling design intended for properly engineered data-center servers. It therefore requires a compatible server chassis with sufficient directed airflow.
3. Complete Technical Specification
| Specification | Details |
|---|---|
| Manufacturer | NVIDIA |
| Product | NVIDIA L40 |
| Product Class | Data Center GPU / Accelerator |
| 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 Interface | 384-bit |
| Memory Bandwidth | 864GB/s |
| FP32 Performance | Up to 90.5 TFLOPS |
| FP16 Tensor Performance | Up to 181 TFLOPS* |
| BF16 Tensor Performance | Up to 181 TFLOPS* |
| FP8 Tensor Performance | Up to 362 TFLOPS* |
| RT Core Performance | Up to 210.6 TFLOPS |
| System Interface | PCIe 4.0 x16 |
| Form Factor | Full-height, full-length, dual-slot |
| Cooling | Passive |
| Maximum Board Power | 300W |
| Power Connector | 16-pin |
| Video Encoders | 2 × NVENC |
| Video Decoders | 2 × NVDEC |
| AV1 Encoding | Supported |
| AV1 Decoding | Supported |
| MIG | Supported |
| Virtualization | NVIDIA vGPU / RTX Virtual Workstation |
| Primary Deployment | Data Center / Enterprise Server |
| Primary Workloads | AI, inference, rendering, visualization, simulation |
| Display Connectivity | Intended for server/data-center deployment rather than conventional desktop display use |
* Tensor performance depends on precision mode and NVIDIA’s specified performance methodology.
4. Applications
The NVIDIA L40 is suitable for:
- Artificial intelligence
- Machine learning
- Deep-learning inference
- Generative AI
- Large-model inference
- Computer vision
- Natural language processing
- Recommendation systems
- 3D rendering
- Ray-traced rendering
- Professional visualization
- NVIDIA Omniverse
- Digital twins
- Engineering visualization
- Scientific visualization
- Simulation
- Virtual workstations
- Remote visualization
- Video processing
- AV1 media workloads
- Virtual production
- Enterprise GPU computing
- Cloud GPU infrastructure
- Data-center graphics acceleration
5. Compatibility
Hardware Compatibility
The NVIDIA L40 is designed for qualified data-center servers supporting:
- PCIe 4.0 x16
- Full-height/full-length dual-slot GPU installation
- 16-pin power connection
- Up to 300W GPU power
- Adequate directed server airflow
- Compatible server BIOS and firmware
- Chassis cooling designed for passive GPUs
Cooling Requirement
The L40 is passively cooled. Unlike a conventional desktop graphics card, it relies on the host server’s airflow system to remove heat.
It should therefore be installed only in a server platform specifically designed to provide sufficient airflow across passive accelerator cards.
Software Compatibility
The L40 supports NVIDIA’s professional and enterprise software ecosystem, including:
- CUDA
- NVIDIA AI Enterprise
- NVIDIA TensorRT
- NVIDIA Triton Inference Server
- NVIDIA NGC
- NVIDIA vGPU
- NVIDIA RTX Virtual Workstation
- NVIDIA Omniverse
- CUDA-accelerated applications
- Professional visualization applications
6. Key Benefits
- 48GB ECC GDDR6 memory
- 864GB/s memory bandwidth
- 18,176 CUDA cores
- 568 fourth-generation Tensor Cores
- 142 third-generation RT Cores
- Up to 90.5 TFLOPS FP32
- Advanced Tensor Core AI acceleration
- Hardware-accelerated ray tracing
- FP8 AI processing support
- AV1 encoding and decoding
- Two NVENC encoders
- Two NVDEC decoders
- PCIe 4.0 x16 interface
- Passive data-center cooling
- 300W power envelope
- Designed for continuous enterprise workloads
- NVIDIA AI Enterprise support
- NVIDIA vGPU support
- Large 48GB GPU memory capacity
- Suitable for multi-GPU server deployments
- Combines AI, graphics, visualization, and compute capabilities
- Designed for professional data-center environments
7. SEO Title
NVIDIA L40 48GB ECC GDDR6 Data Center GPU – Ada Lovelace AI, Rendering & Virtualization Accelerator



