| Product SKU | L40S-48GB |
|---|---|
| FP32 | 91.6 teraFLOPS |
| TF32 Tensor Core | 366 teraFLOPS |
| FP16 | 733 teraFLOPS |
| FP8 | 1,466 teraFLOPS |
| RT Core Performance | 212 teraFLOPS |
| Max Power Consumption | 350W |
| GPU Architecture | NVIDIA Ada Lovelace architecture |
| GPU Memory | 48GB GDDR6 with ECC |
| Memory Bandwidth | 864GB/s |
| Interconnect Interface | PCIe Gen4 x16: 64GB/s bidirectional |
| NVIDIA Ada Lovelace Architecture-Based CUDA® Cores | 18,176 |
| NVIDIA Third-Generation RT Cores | 142 |
| NVIDIA Fourth-Generation Tensor Cores | 568 |
| RT Core Performance TFLOPS | 212 |
| FP32 TFLOPS | 91.6 |
| TF32 Tensor Core TFLOPS | 183 I 366 |
| BFLOAT16 Tensor Core TFLOPS | 362.05 I 733 |
| FP16 Tensor Core | 362.05 I 733 |
| FP8 Tensor Core | 733 I 1,466 |
| Peak INT8 Tensor TOPS Peak INT4 Tensor TOPS | 733 I 1,466* 733 I 1,466 |
| Form Factor | 4.4" (H) x 10.5" (L), dual slot |
| Display Ports | 4x DisplayPort 1.4a |
| Power Connector | 16-pin |
| Thermal | Passive |
| Virtual GPU (vGPU) Software Support | Yes |
| vGPU Profiles Supported | See virtual GPU licensing guide |
| NVENC I NVDEC | 3x l 3x (includes AV1 encode and decode) |
| Secure Boot With Root of Trust | Yes |
| NEBS Ready | Level 3 |
| Multi-Instance GPU (MIG) Support | No |
| NVIDIA® NVLink® Support | No |
Data Center GPUs, GPU
NVIDIA L40S 48GB PCIe
The NVIDIA L40S 48GB PCIe is a data-center GPU built on the NVIDIA Ada Lovelace architecture, with 48GB GDDR6 with ECC memory and 864GB/s of memory bandwidth. It combines 1,466 FP8 and 733 FP16 Tensor Core performance with a board power of 350 W in a 4.4″ (H) x 10.5″ (L), dual slot form factor.
Overview
The NVIDIA L40S 48GB PCIe is a data-center GPU built on the NVIDIA Ada Lovelace architecture that serves AI inference, training and 3D graphics from a single PCIe card. It combines 48 GB of GDDR6 with ECC, 864 GB/s of memory bandwidth and 18,176 CUDA cores with fourth-generation Tensor Cores and third-generation RT Cores. Because it fits a standard PCIe Gen4 x16 slot within a 350 W envelope, it is widely deployed in mainstream servers and virtual workstations.
At a glance
- Architecture: NVIDIA Ada Lovelace; 48 GB GDDR6 with ECC at 864 GB/s
- Tensor Core peaks (with sparsity): 1,466 TFLOPS FP8 · 733 TFLOPS FP16/BF16 · 366 TFLOPS TF32
- CUDA cores: 18,176; third-generation RT Cores: 142
- FP32: 91.6 TFLOPS; RT Core: 212 TFLOPS
- Interconnect: PCIe Gen4 x16 (64 GB/s bidirectional)
- Form factor: dual slot; 350 W TDP; 16-pin power connector
- Media: 3x NVENC and 3x NVDEC (including AV1 encode and decode)
- Secure Boot with Root of Trust; NEBS Level 3 ready; no MIG, no NVLink
When to choose the L40S
The L40S suits teams that need one card for both AI and graphics: LLM inference and fine-tuning, generative AI, and virtual desktop or 3D workloads that also need strong rendering. It does not support NVLink or MIG, so for tightly coupled multi-GPU training the H100/H200 class is the better choice.
Platform & compatibility
The L40S is a standard PCIe Gen4 x16 card and drops into most modern servers that can supply the PCIe lanes, a 350 W power budget and the required 16-pin connector. It is supported for virtual GPU (vGPU) software and is NEBS Level 3 ready for telco and edge deployments.
FAQ
Can the L40S be used for training, not just inference? Yes — it is used for both, though large-scale training usually favours the higher-bandwidth SXM platforms.
Does it support NVLink? No. Multi-GPU communication goes over PCIe.
How much memory does it have? 48 GB of GDDR6 with ECC.
Related
- All NVIDIA data-center GPUs
- Data-center GPU buying guide: A100 to B300
- Browse all GPUs and accelerators
Configuration, availability and lead time depend on the required quantity, server pairing and delivery destination. Send your requirement to receive a quotation.
