NVIDIA hinted that if given the chance, they might launch a standalone version of Arm-based Grace CPU for the server market. To briefly review, this advanced computing chip combines GraceCPU and HopperGPU chips to take full advantage of the entire artificial intelligence and data center ecosystem. NVIDIA hopes to provide customers with a complete set of solutions to focus on different types of workloads in the industry by integrating CPU+GPU into a single design.
Grace uses a dual-chip package and has two products: GPU+CPU and pure CPU. However, this limits customers somewhat, as they are now "locked" into having to use the Hopper architecture or two CPUs, when they may only need one chip. However, Nvidia does have plans to offer its high-end Grace CPUs separately.
Nvidia's current GPU roadmap still shows the company looking to expand its superchip designs into CPU+GPU suites with upcoming architectures such as the GH200, GB200, and GX200, but we may see at the upcoming GTC that the company is bringing Grace CPU or its successor into the standalone server space.
When Nvidia CFO Colette Kress spoke at a Wells Fargo event, she was asked if the company had plans to provide the server industry with the Grace architecture in standalone CPUs. Her answer was clear. Here's what she said:
Will subsequent versions such as GraceHopper, GH200, and GH300 be launched? Or is there just one Grace? Is there a market for Nvidia's ARM-based CPUs?
It is possible to launch only Grace. We may see new product solutions in the data center, and we may also see opportunities for Grace.
-Colette Kress (Chief Financial Officer of NVIDIA)
Although NVIDIA's chief financial officer has not officially confirmed it, judging from the future development of the industry, the company seems to have plans to provide GraceCPU separately. In terms of GraceCPU performance, it has a total of 144 cores (72 Arm NeoverseV2 per chip), supports 960GB of LPDDR5X memory, has a raw bandwidth of 1TB/s, and a comprehensive power consumption of 500W. Other specifications include 117MB L3 cache and 58 Gen5 channels, all using TSMC's 4N process node.
Given that NVIDIA's artificial intelligence products have penetrated deeply into the market, providing GraceCPU in a stand-alone package may change the way data centers operate. In addition, this also means that x86 solutions from companies such as Intel and Nvidia will face more competition, as official benchmarks show that the efficiency and performance of GraceCPU are very competitive with similar products in the industry. Servers are not the only market where Nvidia plans to launch independent CPUs. The company expects to launch Arm-based solutions by 2025 to enter the consumer PC market.