At NVIDIA GTC 2022 Fall, we had several new GPUs launched, and not just the RTX 4000 series. NVIDIA is already looking to bring its GPUs to the next generation in workstations and servers. One of the offerings is an update to the NVIDIA A40 48GB GPU and is called the L40.
NVIDIA L40 Omniverse GPU Launched
NVIDIA did not give many details. Instead, we got a fairly meager spec sheet with minimal details:
From this, there is not a lot that we know. We wish NVIDIA just provided specs of its new cards when they are announced and on its website. We see it is a 48GB with ECC GPU, like the A40. We also see that it is a passive 300W card, like the A40. There is also AV1 encode/ decode. For those who are confused about the display connectors, the A40 had these as well, but only three DisplayPort connectors:
Often with the A40 GPUs, these connectors are not used because these cards are designed for servers. With chassis fans, these are not designed to be desktop GPUs. Instead, NVIDIA is suggesting that the L40 is supposed to be an Omniverse GPU for creative collaboration. This would mean that we expect configurations like the ASUS ESC8000A-E11 8x GPU AMD EPYC server we reviewed.
The Omniverse push for creative platforms backed by cloud access and storage is a big one for AMD, so we expect the NVIDIA L40 to be popular in this generation.
The NVIDIA L40 launch sets up an interesting proposition. The data center GPU will have twice the memory of the desktop RTX 4090, ECC support, and passive cooling, but a much lower TDP than the desktop offering so it can be packed densely into PCIe GPU servers. At the same time, it means we will have “H” series “Hopper” GPUs like the NVIDIA H100 in the data center alongside “L” series Ada Lovelace GPUs.
It feels a bit like NVIDIA is splitting its GPUs into those that are primarily designed to accelerate graphics (“L”) and those that are more designed for maximum AI and HPC compute (“H”), so we will see how the product lines mix as we move forward.