Used NVIDIA L40S: Price, Listings & Specs | BreezeWish

The NVIDIA L40S is an Ada Lovelace PCIe card aimed at inference, graphics, and mixed visual workloads. NVIDIA lists 48 GB of GDDR6 with ECC, 864 GB/s of memory bandwidth, and a 350 W thermal design power. The link is PCIe Gen4 x16. There is no NVLink on this card [S1]. NVIDIA announced the L40S in 2023 as a universal data center GPU for inference and graphics [S2]. Cards show up for sale as single GPUs or in small servers, and the price is the seller's.

Memory is 48 GB of GDDR6 with ECC, not HBM, and the card stands alone in the slot [S1]. Buyers use it where a full HGX pod is more than they need, or where a server already has PCIe slots and 350 W of power and airflow to spare. Because there is no NVLink, adding cards scales over PCIe and the network. Ask how many cards the chassis actually ran, whether the power supply and risers are included, and whether the lot is cards only or a running node. Confirm the report says L40S and 48 GB, not an L40 or another Ada PCIe card. Sellers should set the model to L40S, the form factor to PCIe, and memory to 48, and should say if the original bracket and power cables ship with the cards. A box of cards without the server is a different purchase from a running inference node.