NVIDIA made an interesting announcement at SIGGRAPH this week, introducing a new CPU and GPU combination that indicates a change in its business approach. The company unveiled a next-generation GH200 targeting 2024, but with a twist: it demoted Intel and AMD CPUs in favor of its own future H100 GPU.

The highlight of NVIDIA’s announcement is the new “dual configuration” Grace Hopper GH200, featuring an updated GPU component. Instead of using the current HBM3 memory, NVIDIA will utilize an HBM3e-based Hopper architecture.

The current NVIDIA H100 offers 80GB of HBM3 memory, achieved through five active 16GB HBM3 stacks.

The next-gen GH200, according to NVIDIA’s announcement, comprises a single server equipped with 144 Arm Neoverse cores, delivering up to 3.5 times more memory capacity and 3 times more bandwidth than the current generation. It also features eight petaflops of AI performance and 282GB of the latest HBM3e memory technology.

The 144 Arm cores are easily identifiable in NVIDIA’s accompanying image, showing two Grace 72-core Arm CPUs in the dual configuration.

Dividing the figures in half gives us two Grace CPUs with 72 cores each. As for the 282GB of HBM3e memory, if we assume that NVIDIA is using 24GB stacks instead of 16GB ones, we can calculate that there are six HBM3e spots, totaling 144GB. This configuration closely aligns with what NVIDIA is likely planning for the H100 side of the next-gen Grace Hopper.

This business strategy of refreshing their GPUs with increased memory capacity is not new for NVIDIA. In the past, the company has released mid-cycle refreshes with larger memory options for their GPUs.

The significance of this announcement lies in the fact that NVIDIA is introducing its flagship GPU first on its own Arm CPU platform. This suggests that NVIDIA is aiming to strengthen and favor its own CPUs and ecosystem over others. With the demand and pricing power it currently enjoys, NVIDIA has the ability to offer superior versions of its GPUs exclusively on its own CPUs, which can command a significant premium.

While this may be advantageous for NVIDIA, it also creates opportunities for competitors to explore alternative architectures and potentially gain market share as margins increase.