Nvidia Updates Chip Roadmap to Gear up for AI and Gaming

Nvidia provided a peek into its hardware future as it tries to solidify its dominance in systems that run artificial intelligence and gaming applications.
The company made a name with its GPUs for gaming, but its chips are now in demand for enterprise applications. At the Computex trade show this week, the company introduced new hardware with homegrown CPUs and GPUs for cloud gaming, high-performance computing, metaverse and artificial intelligence.
Grace CPUs
The four systems have Nvidia’s GPUs and homegrown CPUs called Grace, which is based on ARM architecture. Nvidia earlier this year walked away from its plan to acquire ARM for $40 billion after opposition from governments in the EU, US, UK and China.
“These new reference designs will enable our ecosystem to rapidly productize the servers that are optimized for Nvidia-accelerated computing software stacks, and can be qualified as a part of our Nvidia certified systems lineup,” said Paresh Kharya, senior director of product management for accelerated computing at Nvidia, during a press briefing.
The CGX reference board for cloud graphics and gaming has 16 GPUs and what Nvidia calls a Grace “superchip,” which has two CPUs connected by a superfast NVLink interconnect. The superchip has 144 CPU cores, LPDDR5x memory and memory bandwidth of 1 terabyte per second.
The OVX hardware has CPUs, GPUs and an integrated chip called Bluefield, which also packages networking components to process high volumes of data in a small footprint. The OVX is designed for the creation of “digital twins,” which involves creating a virtual version of real-world objects. The OVX is designed for metaverse-style applications in which engineers work together to virtually design and test products like planes, engines, and autonomous cars.
Nvidia also announced HGX servers with its Grace CPUs and GPUs based on the older Ampere architecture. Newer HGX servers with GPUs based on the new Hopper architecture will ship toward the end of this year.
Nvidia’s servers are optimized to work with the CUDA parallel programming framework, which takes advantage of the company’s unique GPU features. The servers will also work with other parallel programming frameworks that include OpenCL and Intel’s OneAPI, but those will not deliver the same performance as CUDA.
The new hardware puts Nvidia’s Grace CPU in competition with Intel’s x86 server chips, which dominate systems today. Nvidia’s GPUs are paired mostly with Intel’s x86 CPUs in high-performance systems, but the company isn’t abandoning x86, Kharya said.
“Absolutely, x86 is a very important CPU — it’s pretty much all of the markets of Nvidia’s GPUs today and we will continue to support x86, will continue to support Arm-based CPUs, offering our customers the choice for wherever they want to deploy accelerated computing,” Kharya said.
Full-Stack Provider
Nvidia is repositioning itself as a full-stack provider of hardware, software and services for AI and graphics applications. The company has software called Enterprise AI in which it packages software stacks targeted at verticals that include health care and automotive.
The software suites include data preparation, analysis and inferencing tools required by developers and data scientists for AI deployment. For example, Nvidia’s Drive platform has a software stack on which carmakers can build autonomous hardware systems, but it requires Nvidia’s GPUs.
The company sees a $1 trillion market opportunity around software, including $150 billion for AI simulation and training, $150 billion for digital twins and $100 billion in gaming.