Nvidia flexes generative AI muscle at SIGGRAPH with new GPUs, development software

Looking to solidify its position as the dominant global supplier of chips that support generative AI workoads, Nvidia announced new GPUs and servers as well as a range of new software offerings at the SIGGRAPH conference in Los Angeles this week.

On the hardware side, Nvidia announced a new line of servers, the OVX series. The server line is designed to use up to eight of the company’s L40S GPUs. The GPUs are based on the company’s Ada Lovelace architecture, which succeeded Ampere as the microarchitecture in use in its main line graphics cards. Each L40S packs 48GB of memory and is designed with complex AI workloads in mind, boasting 1.45 petaflops of tensor processing power.

It’s similar to the approach Nvidia has taken in the past with its consumer graphics card designs, in that the company plans to offer OVX server reference designs, and other manufacturers (in this case, Dell, ASUS, Gigabyte, HPE, Lenovo, QCT and Supermicro) will serve as global system builders. The L40S will become available in the fall, and the company said that OVX systems will go on sale soon after.

As part of an upgrade to its AI Enterprise software line, Nvidia also released a new product called AI Workbench, which is designed to be a sort of self-assembly kit for AI developers. The system comes with pretrained models and an array of tools that can be used to customize them, with the idea of saving considerable development time. Nvidia also announced numerous features designed to add generative AI capabilities to its other product lines,including an AI developer “co-pilot” for its Omniverse 3D imaging software.

How Nvidia targets different sets of users

Many of the company’s newest AI-related releases are targeted at different users — including cloud service providers, developers, and server makers. That’s a key part of Nvidia’s strategy, according to Shane Rau, research vice president at IDC.

“If the end customer’s a cloud service provider, they may just want, say, a server GPU board,” he said. “Some customers would like to buy the Nvidia silicon but also buy the whole system around it — LVX, OVX, and so on. Then maybe the next level is you buy the hardware but maybe you also need some training.”

Another important strategic point, according to Rau, is Nvidia’s flexibility. That flexibility started as long ago as 2012, when the company released its first server GPUs, with the CUDA developer environment that allowed them to be reprogrammed and optimized for different tasks, and has continued with the various AI-related pieces of software that Nvidia has released. The only place, in fact, where the company tends to stop offering solutions is when it would encroach directly on an end user’s own domain.

“AI can be very end-user specific,” Rau said. “Usually the end user brings in their own expertise — agriculture, financial analysis, and so on. So Nvidia wanst to bring the level of solution that you’re wiling to invest in all the way up to your specific domain, but you provide the specific expertise.”

It’s been a highly successful strategy for the company in the AI market, Rau added, given that Nvidia is the largest provider of silicon for AI use by some distance.

“I’d say this was always in the cards for them,” he said.

(Editor’s note: This story has been corrected to clarify that Nvidia will be offering server reference designs, not selling its own branded servers.)