NVIDIA’s SIGGRAPH keynote not only covered things about Omniverse, OpenUSD, and all the AI-related topics, let’s go through the GPU department in this post.
First, we have the new model with a slightly different naming scheme – the NVIDIA L40S. Created for NVIDIA OVX servers, this card is a powerful solution for data centers to process tons of complex applications like AI training and inferencing on top of 3D designs, visualizations, video processing, and more.
Each card contains 48GB of memory built on top of the Ada Lovelace architecture packing 4th Gen Tensor Cores and FP8 Transformer Engine, making up to 1.45 Petaflops of tensor processing power. And the crazy thing is NVIDIA OVX systems can host up to 8 of these per server.
When compared to the older A100, one can expect up to 20% improvement in generative AI inference performance all the way up to 70% for training performance respectively.
On the other hand, the 142 3rd-gen RT cores make up a total of 212 Teraflops of ray-tracing performance which will hugely benefit artists from any sector seeking photorealistic output.
Computational power is at an all-time high too, landing at 18176 CUDA codes creating an advantage of up to 5x higher FP32 performance than the old king A100.
Enterprises that deploy the L40S will be able to benefit from NVIDIA AI Enterprise software right off the bat while those seeking cloud solutions may shift their focus onto CoreWeave. Omniverse Cloud will also support NVIDIA OVX systems for next-level supercharged AI pipeline and workload processing power.
The L40S is slated for release this fall.
As for privately owned businesses that want industrial-class graphic cards, the 3 new members of the Ada Generation GPUs are going to fill that role. Coming in flavors of RTX 5000, RTX 4500, and RTX 4000, they are given GDDR6 VRAM that comes in at 32GB/24GB/20GB respectively with error-correction alongside the signature Ada Lovelace features like 4th Gen Tensor Cores + 3rd Gen RT Cores + DLSS 3.
These can be acquired more or less as a single unit in case users want to customize their own systems according to their needs.
For those seeking a full-package solution and perhaps even stronger performance, OEM partners like HP/Lenovo/BOXX are going to offer their own systems running RTX 6000 GPUs with NVIDIA AI Enterprise and Omniverse Enterprise pre-installed for plug-and-play deployments.
Availability-wise, the RTX 5000 is currently available in the market for purchase while the RTX 4500, RTX 4000, and RTX 6000 equipped systems will come in the following months.
