NVIDIA Announces L40S – New All-In-One Ada Lovelace-Based Accelerator
NVIDIA Corporation has updated a series of L40 accelerators, presented last fall as part of the OVX platform. The new product called NVIDIA L40S is positioned as a universal accelerator in the form factor of a two-slot FHFL expansion card with a PCIe 4.0 x16 interface, suitable for solving almost any task.
In many ways, the L40S repeats the L40 – it is also based on the Ada Lovelace architecture, equipped with an AD102 GPU, supplemented by 48 GB of GDDR6 ECC memory (384 bits, 864 GB / s). The accelerator has 18,176 CUDA cores, 142 third-generation RT cores, and 568 fourth-generation tensor cores. That is, there are no differences from the L40 in this. But the TDP value of the novelty is 50 W higher and is 350 W, it still has passive cooling.
At the same time, L40S manages to be almost twice as fast as L40 in all computing formats using tensor cores, but without Tensor Core, its FP32 performance increased minimally – from 90.5 to 91.6 Tflops. The novelty has not acquired support for the NVLink bridge. The L40S is equipped with four DP 1.4a ports supporting NVIDIA Mosaic and Quadro Sync. vGPU profiles for vDWS, GRID vApps/vPC, vCS are also available. There is support for Secure Boot with Root of Trust and compliance with the NEBS Level 3 standard.
