The NVIDIA L40 is an Ada Lovelace architecture datacenter GPU designed for visual computing, rendering, and AI workloads, announced in September 2022. It represents the successor to the A40, bringing Ada Lovelace architecture improvements to datacenter visualization and inference applications.
The L40 uses the AD102 die with 48GB of GDDR6 ECC memory providing 864 GB/s bandwidth. It includes 18,176 CUDA cores, 568 fourth-generation Tensor Cores, and 142 third-generation RT cores for hardware-accelerated ray tracing. The chip is manufactured on TSMC's custom 4N process.
Unlike the L40S (inference-optimized variant), the L40 emphasizes graphics capabilities including full RT core enablement for professional rendering workloads. The Tensor Cores support standard Ada precision modes (FP16, BF16, TF32, INT8) but not the enhanced FP8 Transformer Engine found in L40S.
The dual-slot PCIe Gen4 x16 form factor has a 300W TDP, enabling air-cooled deployment in standard servers. Hardware video capabilities include 8th-generation NVENC with AV1 encode support and 5th-generation NVDEC. Primary deployment scenarios include professional visualization servers, rendering farms, VDI infrastructure, and cloud gaming platforms.