Nvidia has today at the Siggraph four new workstation graphics cards presented: the Nvidia RTX 5000, RTX 4500 and RTX 4000 for workstations, as well as the Nvidia L40S for OVX systems. They should also make the AI performance of the Nvidia Ada generation available in various performance classes in workstations with extended memory requirements.
What the RTX models have in common is the ability to connect them to a cluster via NV-Link. Manufacturers can offer systems with up to four GPUs, which means a total of 192 GB of graphics memory and 5,828 Tflops of AI computing power in the highest configuration with 48 GB of GDDR6X per GPU. Nvidia mentions them almost exclusively in connection with AI development and Omniverse.
The new GPUs come alongside a new version of the Nvidia Omniverse. It is developed using the Universal Design Language, which now has new cloud APIs. The new functions can be operated both on local systems such as those mentioned above and on OVX servers as a cloud solution.
The L40 will be even faster
It uses the new Nvidia L40S GPU. The L40S is a more powerful offshoot of the already available Nvidia L40. It closely resembles the Nvidia RTX 6000 (Ada) for desktop workstations. For operation in OVX servers, it is equipped with a passive heat sink and relies on a constant flow of air in the server housing. This type of cooling is normal for air-cooled servers with computational accelerators.
Nvidia specifies the power consumption as 350 watts, 50 watts more than the Nvidia L40. The GPU itself is an AD-102 full build with 18,176 FP32 ALUs and 142 RT cores of the 3rd generation. Nvidia states 91.6 Tflops FP32 performance, almost identical to the Nvidia RTX 6000 with 91.1 Tflops. The L40S should be available in the fourth quarter of 2023 in servers from Asus, Dell, Gigabyte, HPE, Lenovo, QCT and Supermicro.