Rent NVIDIA L40S GPUs

The NVIDIA L40S is a data center GPU designed for AI inference, model training and fine-tuning, deep learning and data science work, virtual workstations, 3D design and rendering, and multi-tenant cloud deployments. It features 48GB GDDR6 ECC memory, 733 TFLOPS of FP8 Tensor performance, vGPU support, and enterprise-grade reliability.

The L40S and L40 share the same Ada Lovelace silicon: 18,176 CUDA cores, 568 4th-gen Tensor cores, 142 3rd-gen RT cores, and 48GB of GDDR6 ECC memory at 864 GB/s. The L40S is tuned for AI rather than graphics: it runs at higher clocks with a 350W TDP (versus 300W) and delivers roughly double the Tensor throughput, at 362 TFLOPS BF16/FP16 and 733 TFLOPS FP8 against the L40's 181 and 362. Choose the L40S for inference and fine-tuning; the L40 was positioned more for visualization and vGPU density.

The NVIDIA L40S has 48 GB of GDDR6 memory with ECC (Error-Correcting Code) on a 384-bit bus at 864 GB/s. That is enough to serve most 70B-parameter models quantized to FP8, or to fine-tune mid-size models without sharding across GPUs.

The A100 80GB has more memory (80GB vs 48GB) and far more bandwidth (2,039 GB/s vs 864 GB/s), which makes it stronger for large-model training and memory-bound workloads. The L40S wins on FP8, a format the Ampere-based A100 does not support at all, and adds RT cores plus 3× NVENC/NVDEC with AV1 for rendering and video pipelines. For FP8 inference and mixed graphics/AI workloads the L40S is usually the better value; for large-scale training the A100 is the stronger choice.

Yes, the L40S supports NVIDIA vGPU technology for virtual workstations (vWS) and virtual PCs/apps, making it ideal for multi-tenant cloud environments and remote graphics workloads. Note that unlike the A100 it does not support MIG, so partitioning is done through vGPU rather than hardware instance slicing.

Key features include 18,176 CUDA cores, 568 fourth-generation Tensor Cores, 142 third-generation RT Cores, 48GB GDDR6 ECC memory, 864 GB/s bandwidth, 91.6 TFLOPS FP32, 733 TFLOPS FP8 Tensor, PCIe 4.0 x16, a passive dual-slot 350W design, 3x NVENC/NVDEC with AV1 support, and secure boot with a hardware root of trust for enterprise data center operations.

Yes, on demand at $0.63 per GPU-hour, billed by the second, in instances of up to eight GPUs. Reserved terms bring it to $0.57 for one month and $0.54 for three months. Check the Console for current availability, or contact us to reserve capacity.