At the annual Arm Everywhere China event on September 8, Arm launched the new-generation data center compute subsystem Neoverse CSS N4 and shared the latest ecosystem progress for Arm AGI CPU. Arm also announced deepened cooperation with New Unisplendour Group and plans to establish a joint artificial general intelligence innovation center in China. Against the backdrop of GPUs long dominating AI conversations, the focus of this launch is a more "traditional" role—the CPU.
Arm's reasoning relates to how AI agents operate. As AI applications shift from generating content to executing tasks, an agent often needs to retrieve data, access databases, call external tools, and collaborate with multiple models or other agents beyond completing inference. These operations are not necessarily best handled entirely by GPUs, which increases data center demand for general-purpose CPUs, memory bandwidth, and high-speed I/O.
As a highly configurable compute subsystem, CSS N4 can be configured with 8 to 128 CPU cores per die, at frequencies up to 3.8GHz, and is the first Neoverse N-series CSS to support LPDDR6 memory and PCIe 7.0, with up to 128 PCIe lanes. Arm says that compared with the previous-generation CSS N3, single-socket performance is up to 2x higher, performance per watt is up to 25% higher, and memory bandwidth is up to 75% higher.
These metrics map directly to the "work outside the model" created by agents. Agent sandboxes isolate code and tool calls, DPUs handle networking and data movement, and KV cache stores intermediate inference states. As the number of agents and concurrent tasks grows, these computing workloads—previously situated around GPUs—could further increase consumption of CPU, memory, and network