Product
NVIDIA L40S 48GB
Part number NVIDIA L40S 48GB
- Price
- USD $100
- Stock
- 1
Ada Lovelace data-center GPU with 48 GB of GDDR6 ECC memory, a 350 W passive dual-slot PCIe Gen4 card, and no NVLink, aimed at inference, graphics, and mixed enterprise workloads.
- 48 GB GDDR6 memory with ECC, 864 GB/s
- PCIe Gen4 x16 dual-slot card, 350 W, passive cooling
- Fourth-generation Tensor Cores and third-generation RT Cores
- No NVLink and no Multi-Instance GPU, per NVIDIA’s specification table
- AV1 encode and decode, with vGPU software support
Specifications
| GPU Model | NVIDIA L40S 48GB |
|---|---|
| Architecture | NVIDIA Ada Lovelace |
| Form Factor | PCIe Gen4 dual-slot |
| GPU Count | 1 |
| GPU Memory | 48 GB GDDR6 ECC |
| Memory Bandwidth | 864 GB/s |
| Interconnect | PCIe Gen4 x16, 64 GB/s bidirectional; NVLink not supported |
| Max Power | 350 W |
| Cooling | Passive, dual-slot |
| Mechanical | 4.4 in (H) × 10.5 in (L), dual slot |
| CUDA Cores | 18,176 |
| Tensor Cores | 568 (4th generation) |
| RT Cores | 142 (3rd generation) |
| FP32 | 91.6 TFLOPS |
| RT Core Performance | 212 TFLOPS |
| TF32 Tensor Core | 183 / 366 TFLOPS (with sparsity) |
| FP16 / BFLOAT16 Tensor Core | 362.05 / 733 TFLOPS (with sparsity) |
| FP8 Tensor Core | 733 / 1,466 TFLOPS (with sparsity) |
| INT8 Tensor Core | 733 / 1,466 TOPS (with sparsity) |
| Power Connector | 16-pin |
| Display Ports | 4 × DisplayPort 1.4a |
| NVENC / NVDEC | 3× / 3×, including AV1 |
| Multi-Instance GPU | Not supported |
| NVIDIA NVLink | Not supported |
| vGPU Software | Supported |
Description
The NVIDIA L40S is a universal data-center GPU based on the Ada Lovelace architecture. It is a full-height, full-length, dual-slot PCIe Gen4 card with 48 GB of GDDR6 memory with ECC and 864 GB/s of memory bandwidth. NVIDIA rates maximum power at 350 W and specifies a passive thermal solution and a 16-pin power connector.
Unlike the Hopper and Blackwell SXM modules in this catalog, L40S does not support NVLink or Multi-Instance GPU. The host interface is PCIe Gen4 x16 at 64 GB/s bidirectional. What it does offer is a broad mix of compute and graphics hardware: 18,176 CUDA cores, 568 fourth-generation Tensor Cores, 142 third-generation RT Cores, and 91.6 TFLOPS of FP32. With sparsity, NVIDIA publishes 733 TFLOPS of FP16/BF16 Tensor performance and 1,466 TFLOPS of FP8 Tensor performance. Three NVENC and three NVDEC engines include AV1, and the card has four DisplayPort 1.4a outputs. NVIDIA also lists vGPU software support, Secure Boot with a root of trust, and NEBS Level 3.
L40S is typically used for inference, multimodal generative workloads, rendering, virtual workstations, and mixed enterprise graphics-and-compute nodes — including NVIDIA OVX server designs. It is not a substitute for an eight-GPU HGX training board. A verified four-GPU host is the GIGABYTE G293-S47, which Gigabyte documents as optimized for an NVIDIA OVX L40S four-GPU configuration. Request a quote for the card or a four-GPU inference server.
Related products
GIGABYTE G293-S47-AAP1
P/N G293-S47-AAP1
Price USD $100 Stock 1
2U dual-socket Intel Xeon server optimized for a four-GPU NVIDIA OVX L40S configuration, with four FHFL PCIe Gen5 x16 GPU slots, dual 3000 W Titanium power supplies, and an operating range of 10°C to 25°C.
- Model
- G293-S47-AAP1
- GPU Model
- NVIDIA L40S (OVX 4-GPU configuration)
- GPU Count
- 4 PCIe
- Form Factor
- 2U rackmount
Ready to specify a configuration?
Share the use case, the target build, quantity, destination country, and timeline. We will take it from the inquiry.