NVIDIA L40S vs L4

What Is the Main Difference Between NVIDIA L40S and NVIDIA L4?

L40S and L4 share the Ada Lovelace architecture and fourth-generation Tensor Cores, but they target different performance envelopes.

L40S doubles memory capacity to 48 GB versus 24 GB on L4, increases bandwidth from 300 GB/s to 864 GB/s and lists roughly three times the FP32 and sparse FP8 Tensor Core throughput. It also carries a much higher 350 W maximum power rating.

L4's strength is efficiency: 72 W, single-slot low-profile deployment and strong media support. L40S is the higher-performance choice for larger inference models, rendering and heavier generative AI.

L40S vs L4 for LLM Inference

Both support Ada FP8, but L40S has twice the memory and far more Tensor Core throughput. That makes it better for larger models, bigger batches and higher per-GPU throughput. L4 is well suited to smaller or quantized models where 24 GB is enough.

L40S vs L4 for Graphics and Generative AI

Both include Ada RT Cores, but L40S is the substantially more powerful graphics GPU and carries 48 GB of memory. It is better suited to rendering, virtual workstations, 3D workflows and heavier image-generation workloads.

L40S vs L4 for Power and Server Density

L4 uses only 72 W and fits a low-profile single-slot PCIe form factor. L40S can consume up to 350 W and uses a dual-slot card. For dense inference fleets, L4 can therefore be operationally easier even when L40S is much faster per GPU.